lib/crunchrun/crunchrun.go

   1 // Copyright (C) The Arvados Authors. All rights reserved.
   2 //
   3 // SPDX-License-Identifier: AGPL-3.0
   4
   5 package crunchrun
   6
   7 import (
   8         "bytes"
   9         "context"
  10         "encoding/json"
  11         "errors"
  12         "flag"
  13         "fmt"
  14         "io"
  15         "io/fs"
  16         "io/ioutil"
  17         "log"
  18         "net"
  19         "net/http"
  20         "os"
  21         "os/exec"
  22         "os/signal"
  23         "os/user"
  24         "path"
  25         "path/filepath"
  26         "regexp"
  27         "runtime"
  28         "runtime/pprof"
  29         "sort"
  30         "strings"
  31         "sync"
  32         "syscall"
  33         "time"
  34
  35         "git.arvados.org/arvados.git/lib/cloud"
  36         "git.arvados.org/arvados.git/lib/cmd"
  37         "git.arvados.org/arvados.git/lib/config"
  38         "git.arvados.org/arvados.git/lib/crunchstat"
  39         "git.arvados.org/arvados.git/sdk/go/arvados"
  40         "git.arvados.org/arvados.git/sdk/go/arvadosclient"
  41         "git.arvados.org/arvados.git/sdk/go/ctxlog"
  42         "git.arvados.org/arvados.git/sdk/go/keepclient"
  43         "git.arvados.org/arvados.git/sdk/go/manifest"
  44         "golang.org/x/sys/unix"
  45 )
  46
  47 type command struct{}
  48
  49 var Command = command{}
  50
  51 // ConfigData contains environment variables and (when needed) cluster
  52 // configuration, passed from dispatchcloud to crunch-run on stdin.
  53 type ConfigData struct {
  54         Env          map[string]string
  55         KeepBuffers  int
  56         EC2SpotCheck bool
  57         Cluster      *arvados.Cluster
  58 }
  59
  60 // IArvadosClient is the minimal Arvados API methods used by crunch-run.
  61 type IArvadosClient interface {
  62         Create(resourceType string, parameters arvadosclient.Dict, output interface{}) error
  63         Get(resourceType string, uuid string, parameters arvadosclient.Dict, output interface{}) error
  64         Update(resourceType string, uuid string, parameters arvadosclient.Dict, output interface{}) error
  65         Call(method, resourceType, uuid, action string, parameters arvadosclient.Dict, output interface{}) error
  66         CallRaw(method string, resourceType string, uuid string, action string, parameters arvadosclient.Dict) (reader io.ReadCloser, err error)
  67         Discovery(key string) (interface{}, error)
  68 }
  69
  70 // ErrCancelled is the error returned when the container is cancelled.
  71 var ErrCancelled = errors.New("Cancelled")
  72
  73 // IKeepClient is the minimal Keep API methods used by crunch-run.
  74 type IKeepClient interface {
  75         BlockWrite(context.Context, arvados.BlockWriteOptions) (arvados.BlockWriteResponse, error)
  76         ReadAt(locator string, p []byte, off int) (int, error)
  77         ManifestFileReader(m manifest.Manifest, filename string) (arvados.File, error)
  78         LocalLocator(locator string) (string, error)
  79         ClearBlockCache()
  80         SetStorageClasses(sc []string)
  81 }
  82
  83 // NewLogWriter is a factory function to create a new log writer.
  84 type NewLogWriter func(name string) (io.WriteCloser, error)
  85
  86 type RunArvMount func(cmdline []string, tok string) (*exec.Cmd, error)
  87
  88 type MkTempDir func(string, string) (string, error)
  89
  90 type PsProcess interface {
  91         CmdlineSlice() ([]string, error)
  92 }
  93
  94 // ContainerRunner is the main stateful struct used for a single execution of a
  95 // container.
  96 type ContainerRunner struct {
  97         executor       containerExecutor
  98         executorStdin  io.Closer
  99         executorStdout io.Closer
 100         executorStderr io.Closer
 101
 102         // Dispatcher client is initialized with the Dispatcher token.
 103         // This is a privileged token used to manage container status
 104         // and logs.
 105         //
 106         // We have both dispatcherClient and DispatcherArvClient
 107         // because there are two different incompatible Arvados Go
 108         // SDKs and we have to use both (hopefully this gets fixed in
 109         // #14467)
 110         dispatcherClient     *arvados.Client
 111         DispatcherArvClient  IArvadosClient
 112         DispatcherKeepClient IKeepClient
 113
 114         // Container client is initialized with the Container token
 115         // This token controls the permissions of the container, and
 116         // must be used for operations such as reading collections.
 117         //
 118         // Same comment as above applies to
 119         // containerClient/ContainerArvClient.
 120         containerClient     *arvados.Client
 121         ContainerArvClient  IArvadosClient
 122         ContainerKeepClient IKeepClient
 123
 124         Container     arvados.Container
 125         token         string
 126         ExitCode      *int
 127         NewLogWriter  NewLogWriter
 128         CrunchLog     *ThrottledLogger
 129         logUUID       string
 130         logMtx        sync.Mutex
 131         LogCollection arvados.CollectionFileSystem
 132         LogsPDH       *string
 133         RunArvMount   RunArvMount
 134         MkTempDir     MkTempDir
 135         ArvMount      *exec.Cmd
 136         ArvMountPoint string
 137         HostOutputDir string
 138         Volumes       map[string]struct{}
 139         OutputPDH     *string
 140         SigChan       chan os.Signal
 141         ArvMountExit  chan error
 142         SecretMounts  map[string]arvados.Mount
 143         MkArvClient   func(token string) (IArvadosClient, IKeepClient, *arvados.Client, error)
 144         finalState    string
 145         parentTemp    string
 146         costStartTime time.Time
 147
 148         keepstore        *exec.Cmd
 149         keepstoreLogger  io.WriteCloser
 150         keepstoreLogbuf  *bufThenWrite
 151         statLogger       io.WriteCloser
 152         statReporter     *crunchstat.Reporter
 153         hoststatLogger   io.WriteCloser
 154         hoststatReporter *crunchstat.Reporter
 155         statInterval     time.Duration
 156         // What we tell docker to use as the container's cgroup
 157         // parent.
 158         setCgroupParent string
 159         // Fake root dir where crunchstat.Reporter should read OS
 160         // files, for testing.
 161         crunchstatFakeFS fs.FS
 162
 163         cStateLock sync.Mutex
 164         cCancelled bool // StopContainer() invoked
 165
 166         enableMemoryLimit bool
 167         enableNetwork     string // one of "default" or "always"
 168         networkMode       string // "none", "host", or "" -- passed through to executor
 169         brokenNodeHook    string // script to run if node appears to be broken
 170         arvMountLog       *ThrottledLogger
 171
 172         containerWatchdogInterval time.Duration
 173
 174         gateway Gateway
 175
 176         prices     []cloud.InstancePrice
 177         pricesLock sync.Mutex
 178 }
 179
 180 // setupSignals sets up signal handling to gracefully terminate the
 181 // underlying container and update state when receiving a TERM, INT or
 182 // QUIT signal.
 183 func (runner *ContainerRunner) setupSignals() {
 184         runner.SigChan = make(chan os.Signal, 1)
 185         signal.Notify(runner.SigChan, syscall.SIGTERM)
 186         signal.Notify(runner.SigChan, syscall.SIGINT)
 187         signal.Notify(runner.SigChan, syscall.SIGQUIT)
 188
 189         go func(sig chan os.Signal) {
 190                 for s := range sig {
 191                         runner.stop(s)
 192                 }
 193         }(runner.SigChan)
 194 }
 195
 196 // stop the underlying container.
 197 func (runner *ContainerRunner) stop(sig os.Signal) {
 198         runner.cStateLock.Lock()
 199         defer runner.cStateLock.Unlock()
 200         if sig != nil {
 201                 runner.CrunchLog.Printf("caught signal: %v", sig)
 202         }
 203         runner.cCancelled = true
 204         runner.CrunchLog.Printf("stopping container")
 205         err := runner.executor.Stop()
 206         if err != nil {
 207                 runner.CrunchLog.Printf("error stopping container: %s", err)
 208         }
 209 }
 210
 211 var errorBlacklist = []string{
 212         "(?ms).*[Cc]annot connect to the Docker daemon.*",
 213         "(?ms).*oci runtime error.*starting container process.*container init.*mounting.*to rootfs.*no such file or directory.*",
 214         "(?ms).*grpc: the connection is unavailable.*",
 215 }
 216
 217 func (runner *ContainerRunner) runBrokenNodeHook() {
 218         if runner.brokenNodeHook == "" {
 219                 path := filepath.Join(lockdir, brokenfile)
 220                 runner.CrunchLog.Printf("Writing %s to mark node as broken", path)
 221                 f, err := os.OpenFile(path, os.O_CREATE|os.O_RDWR, 0700)
 222                 if err != nil {
 223                         runner.CrunchLog.Printf("Error writing %s: %s", path, err)
 224                         return
 225                 }
 226                 f.Close()
 227         } else {
 228                 runner.CrunchLog.Printf("Running broken node hook %q", runner.brokenNodeHook)
 229                 // run killme script
 230                 c := exec.Command(runner.brokenNodeHook)
 231                 c.Stdout = runner.CrunchLog
 232                 c.Stderr = runner.CrunchLog
 233                 err := c.Run()
 234                 if err != nil {
 235                         runner.CrunchLog.Printf("Error running broken node hook: %v", err)
 236                 }
 237         }
 238 }
 239
 240 func (runner *ContainerRunner) checkBrokenNode(goterr error) bool {
 241         for _, d := range errorBlacklist {
 242                 if m, e := regexp.MatchString(d, goterr.Error()); m && e == nil {
 243                         runner.CrunchLog.Printf("Error suggests node is unable to run containers: %v", goterr)
 244                         runner.runBrokenNodeHook()
 245                         return true
 246                 }
 247         }
 248         return false
 249 }
 250
 251 // LoadImage determines the docker image id from the container record and
 252 // checks if it is available in the local Docker image store.  If not, it loads
 253 // the image from Keep.
 254 func (runner *ContainerRunner) LoadImage() (string, error) {
 255         runner.CrunchLog.Printf("Fetching Docker image from collection '%s'", runner.Container.ContainerImage)
 256
 257         d, err := os.Open(runner.ArvMountPoint + "/by_id/" + runner.Container.ContainerImage)
 258         if err != nil {
 259                 return "", err
 260         }
 261         defer d.Close()
 262         allfiles, err := d.Readdirnames(-1)
 263         if err != nil {
 264                 return "", err
 265         }
 266         var tarfiles []string
 267         for _, fnm := range allfiles {
 268                 if strings.HasSuffix(fnm, ".tar") {
 269                         tarfiles = append(tarfiles, fnm)
 270                 }
 271         }
 272         if len(tarfiles) == 0 {
 273                 return "", fmt.Errorf("image collection does not include a .tar image file")
 274         }
 275         if len(tarfiles) > 1 {
 276                 return "", fmt.Errorf("cannot choose from multiple tar files in image collection: %v", tarfiles)
 277         }
 278         imageID := tarfiles[0][:len(tarfiles[0])-4]
 279         imageTarballPath := runner.ArvMountPoint + "/by_id/" + runner.Container.ContainerImage + "/" + imageID + ".tar"
 280         runner.CrunchLog.Printf("Using Docker image id %q", imageID)
 281
 282         runner.CrunchLog.Print("Loading Docker image from keep")
 283         err = runner.executor.LoadImage(imageID, imageTarballPath, runner.Container, runner.ArvMountPoint,
 284                 runner.containerClient)
 285         if err != nil {
 286                 return "", err
 287         }
 288
 289         return imageID, nil
 290 }
 291
 292 func (runner *ContainerRunner) ArvMountCmd(cmdline []string, token string) (c *exec.Cmd, err error) {
 293         c = exec.Command(cmdline[0], cmdline[1:]...)
 294
 295         // Copy our environment, but override ARVADOS_API_TOKEN with
 296         // the container auth token.
 297         c.Env = nil
 298         for _, s := range os.Environ() {
 299                 if !strings.HasPrefix(s, "ARVADOS_API_TOKEN=") {
 300                         c.Env = append(c.Env, s)
 301                 }
 302         }
 303         c.Env = append(c.Env, "ARVADOS_API_TOKEN="+token)
 304
 305         w, err := runner.NewLogWriter("arv-mount")
 306         if err != nil {
 307                 return nil, err
 308         }
 309         runner.arvMountLog = NewThrottledLogger(w)
 310         scanner := logScanner{
 311                 Patterns: []string{
 312                         "Keep write error",
 313                         "Block not found error",
 314                         "Unhandled exception during FUSE operation",
 315                 },
 316                 ReportFunc: func(pattern, text string) {
 317                         runner.updateRuntimeStatus(arvadosclient.Dict{
 318                                 "warning":       "arv-mount: " + pattern,
 319                                 "warningDetail": text,
 320                         })
 321                 },
 322         }
 323         c.Stdout = runner.arvMountLog
 324         c.Stderr = io.MultiWriter(runner.arvMountLog, os.Stderr, &scanner)
 325
 326         runner.CrunchLog.Printf("Running %v", c.Args)
 327
 328         err = c.Start()
 329         if err != nil {
 330                 return nil, err
 331         }
 332
 333         statReadme := make(chan bool)
 334         runner.ArvMountExit = make(chan error)
 335
 336         keepStatting := true
 337         go func() {
 338                 for keepStatting {
 339                         time.Sleep(100 * time.Millisecond)
 340                         _, err = os.Stat(fmt.Sprintf("%s/by_id/README", runner.ArvMountPoint))
 341                         if err == nil {
 342                                 keepStatting = false
 343                                 statReadme <- true
 344                         }
 345                 }
 346                 close(statReadme)
 347         }()
 348
 349         go func() {
 350                 mnterr := c.Wait()
 351                 if mnterr != nil {
 352                         runner.CrunchLog.Printf("Arv-mount exit error: %v", mnterr)
 353                 }
 354                 runner.ArvMountExit <- mnterr
 355                 close(runner.ArvMountExit)
 356         }()
 357
 358         select {
 359         case <-statReadme:
 360                 break
 361         case err := <-runner.ArvMountExit:
 362                 runner.ArvMount = nil
 363                 keepStatting = false
 364                 return nil, err
 365         }
 366
 367         return c, nil
 368 }
 369
 370 func (runner *ContainerRunner) SetupArvMountPoint(prefix string) (err error) {
 371         if runner.ArvMountPoint == "" {
 372                 runner.ArvMountPoint, err = runner.MkTempDir(runner.parentTemp, prefix)
 373         }
 374         return
 375 }
 376
 377 func copyfile(src string, dst string) (err error) {
 378         srcfile, err := os.Open(src)
 379         if err != nil {
 380                 return
 381         }
 382
 383         os.MkdirAll(path.Dir(dst), 0777)
 384
 385         dstfile, err := os.Create(dst)
 386         if err != nil {
 387                 return
 388         }
 389         _, err = io.Copy(dstfile, srcfile)
 390         if err != nil {
 391                 return
 392         }
 393
 394         err = srcfile.Close()
 395         err2 := dstfile.Close()
 396
 397         if err != nil {
 398                 return
 399         }
 400
 401         if err2 != nil {
 402                 return err2
 403         }
 404
 405         return nil
 406 }
 407
 408 func (runner *ContainerRunner) SetupMounts() (map[string]bindmount, error) {
 409         bindmounts := map[string]bindmount{}
 410         err := runner.SetupArvMountPoint("keep")
 411         if err != nil {
 412                 return nil, fmt.Errorf("While creating keep mount temp dir: %v", err)
 413         }
 414
 415         token, err := runner.ContainerToken()
 416         if err != nil {
 417                 return nil, fmt.Errorf("could not get container token: %s", err)
 418         }
 419         runner.CrunchLog.Printf("container token %q", token)
 420
 421         pdhOnly := true
 422         tmpcount := 0
 423         arvMountCmd := []string{
 424                 "arv-mount",
 425                 "--foreground",
 426                 "--read-write",
 427                 "--storage-classes", strings.Join(runner.Container.OutputStorageClasses, ","),
 428                 fmt.Sprintf("--crunchstat-interval=%v", runner.statInterval.Seconds())}
 429
 430         if _, isdocker := runner.executor.(*dockerExecutor); isdocker {
 431                 arvMountCmd = append(arvMountCmd, "--allow-other")
 432         }
 433
 434         if runner.Container.RuntimeConstraints.KeepCacheDisk > 0 {
 435                 keepcachedir, err := runner.MkTempDir(runner.parentTemp, "keepcache")
 436                 if err != nil {
 437                         return nil, fmt.Errorf("while creating keep cache temp dir: %v", err)
 438                 }
 439                 arvMountCmd = append(arvMountCmd, "--disk-cache", "--disk-cache-dir", keepcachedir, "--file-cache", fmt.Sprintf("%d", runner.Container.RuntimeConstraints.KeepCacheDisk))
 440         } else if runner.Container.RuntimeConstraints.KeepCacheRAM > 0 {
 441                 arvMountCmd = append(arvMountCmd, "--ram-cache", "--file-cache", fmt.Sprintf("%d", runner.Container.RuntimeConstraints.KeepCacheRAM))
 442         }
 443
 444         collectionPaths := []string{}
 445         needCertMount := true
 446         type copyFile struct {
 447                 src  string
 448                 bind string
 449         }
 450         var copyFiles []copyFile
 451
 452         var binds []string
 453         for bind := range runner.Container.Mounts {
 454                 binds = append(binds, bind)
 455         }
 456         for bind := range runner.SecretMounts {
 457                 if _, ok := runner.Container.Mounts[bind]; ok {
 458                         return nil, fmt.Errorf("secret mount %q conflicts with regular mount", bind)
 459                 }
 460                 if runner.SecretMounts[bind].Kind != "json" &&
 461                         runner.SecretMounts[bind].Kind != "text" {
 462                         return nil, fmt.Errorf("secret mount %q type is %q but only 'json' and 'text' are permitted",
 463                                 bind, runner.SecretMounts[bind].Kind)
 464                 }
 465                 binds = append(binds, bind)
 466         }
 467         sort.Strings(binds)
 468
 469         for _, bind := range binds {
 470                 mnt, notSecret := runner.Container.Mounts[bind]
 471                 if !notSecret {
 472                         mnt = runner.SecretMounts[bind]
 473                 }
 474                 if bind == "stdout" || bind == "stderr" {
 475                         // Is it a "file" mount kind?
 476                         if mnt.Kind != "file" {
 477                                 return nil, fmt.Errorf("unsupported mount kind '%s' for %s: only 'file' is supported", mnt.Kind, bind)
 478                         }
 479
 480                         // Does path start with OutputPath?
 481                         prefix := runner.Container.OutputPath
 482                         if !strings.HasSuffix(prefix, "/") {
 483                                 prefix += "/"
 484                         }
 485                         if !strings.HasPrefix(mnt.Path, prefix) {
 486                                 return nil, fmt.Errorf("%s path does not start with OutputPath: %s, %s", strings.Title(bind), mnt.Path, prefix)
 487                         }
 488                 }
 489
 490                 if bind == "stdin" {
 491                         // Is it a "collection" mount kind?
 492                         if mnt.Kind != "collection" && mnt.Kind != "json" {
 493                                 return nil, fmt.Errorf("unsupported mount kind '%s' for stdin: only 'collection' and 'json' are supported", mnt.Kind)
 494                         }
 495                 }
 496
 497                 if bind == "/etc/arvados/ca-certificates.crt" {
 498                         needCertMount = false
 499                 }
 500
 501                 if strings.HasPrefix(bind, runner.Container.OutputPath+"/") && bind != runner.Container.OutputPath+"/" {
 502                         if mnt.Kind != "collection" && mnt.Kind != "text" && mnt.Kind != "json" {
 503                                 return nil, fmt.Errorf("only mount points of kind 'collection', 'text' or 'json' are supported underneath the output_path for %q, was %q", bind, mnt.Kind)
 504                         }
 505                 }
 506
 507                 switch {
 508                 case mnt.Kind == "collection" && bind != "stdin":
 509                         var src string
 510                         if mnt.UUID != "" && mnt.PortableDataHash != "" {
 511                                 return nil, fmt.Errorf("cannot specify both 'uuid' and 'portable_data_hash' for a collection mount")
 512                         }
 513                         if mnt.UUID != "" {
 514                                 if mnt.Writable {
 515                                         return nil, fmt.Errorf("writing to existing collections currently not permitted")
 516                                 }
 517                                 pdhOnly = false
 518                                 src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.UUID)
 519                         } else if mnt.PortableDataHash != "" {
 520                                 if mnt.Writable && !strings.HasPrefix(bind, runner.Container.OutputPath+"/") {
 521                                         return nil, fmt.Errorf("can never write to a collection specified by portable data hash")
 522                                 }
 523                                 idx := strings.Index(mnt.PortableDataHash, "/")
 524                                 if idx > 0 {
 525                                         mnt.Path = path.Clean(mnt.PortableDataHash[idx:])
 526                                         mnt.PortableDataHash = mnt.PortableDataHash[0:idx]
 527                                         runner.Container.Mounts[bind] = mnt
 528                                 }
 529                                 src = fmt.Sprintf("%s/by_id/%s", runner.ArvMountPoint, mnt.PortableDataHash)
 530                                 if mnt.Path != "" && mnt.Path != "." {
 531                                         if strings.HasPrefix(mnt.Path, "./") {
 532                                                 mnt.Path = mnt.Path[2:]
 533                                         } else if strings.HasPrefix(mnt.Path, "/") {
 534                                                 mnt.Path = mnt.Path[1:]
 535                                         }
 536                                         src += "/" + mnt.Path
 537                                 }
 538                         } else {
 539                                 src = fmt.Sprintf("%s/tmp%d", runner.ArvMountPoint, tmpcount)
 540                                 arvMountCmd = append(arvMountCmd, "--mount-tmp", fmt.Sprintf("tmp%d", tmpcount))
 541                                 tmpcount++
 542                         }
 543                         if mnt.Writable {
 544                                 if bind == runner.Container.OutputPath {
 545                                         runner.HostOutputDir = src
 546                                         bindmounts[bind] = bindmount{HostPath: src}
 547                                 } else if strings.HasPrefix(bind, runner.Container.OutputPath+"/") {
 548                                         copyFiles = append(copyFiles, copyFile{src, runner.HostOutputDir + bind[len(runner.Container.OutputPath):]})
 549                                 } else {
 550                                         bindmounts[bind] = bindmount{HostPath: src}
 551                                 }
 552                         } else {
 553                                 bindmounts[bind] = bindmount{HostPath: src, ReadOnly: true}
 554                         }
 555                         collectionPaths = append(collectionPaths, src)
 556
 557                 case mnt.Kind == "tmp":
 558                         var tmpdir string
 559                         tmpdir, err = runner.MkTempDir(runner.parentTemp, "tmp")
 560                         if err != nil {
 561                                 return nil, fmt.Errorf("while creating mount temp dir: %v", err)
 562                         }
 563                         st, staterr := os.Stat(tmpdir)
 564                         if staterr != nil {
 565                                 return nil, fmt.Errorf("while Stat on temp dir: %v", staterr)
 566                         }
 567                         err = os.Chmod(tmpdir, st.Mode()|os.ModeSetgid|0777)
 568                         if staterr != nil {
 569                                 return nil, fmt.Errorf("while Chmod temp dir: %v", err)
 570                         }
 571                         bindmounts[bind] = bindmount{HostPath: tmpdir}
 572                         if bind == runner.Container.OutputPath {
 573                                 runner.HostOutputDir = tmpdir
 574                         }
 575
 576                 case mnt.Kind == "json" || mnt.Kind == "text":
 577                         var filedata []byte
 578                         if mnt.Kind == "json" {
 579                                 filedata, err = json.Marshal(mnt.Content)
 580                                 if err != nil {
 581                                         return nil, fmt.Errorf("encoding json data: %v", err)
 582                                 }
 583                         } else {
 584                                 text, ok := mnt.Content.(string)
 585                                 if !ok {
 586                                         return nil, fmt.Errorf("content for mount %q must be a string", bind)
 587                                 }
 588                                 filedata = []byte(text)
 589                         }
 590
 591                         tmpdir, err := runner.MkTempDir(runner.parentTemp, mnt.Kind)
 592                         if err != nil {
 593                                 return nil, fmt.Errorf("creating temp dir: %v", err)
 594                         }
 595                         tmpfn := filepath.Join(tmpdir, "mountdata."+mnt.Kind)
 596                         err = ioutil.WriteFile(tmpfn, filedata, 0444)
 597                         if err != nil {
 598                                 return nil, fmt.Errorf("writing temp file: %v", err)
 599                         }
 600                         if strings.HasPrefix(bind, runner.Container.OutputPath+"/") && (notSecret || runner.Container.Mounts[runner.Container.OutputPath].Kind != "collection") {
 601                                 // In most cases, if the container
 602                                 // specifies a literal file inside the
 603                                 // output path, we copy it into the
 604                                 // output directory (either a mounted
 605                                 // collection or a staging area on the
 606                                 // host fs). If it's a secret, it will
 607                                 // be skipped when copying output from
 608                                 // staging to Keep later.
 609                                 copyFiles = append(copyFiles, copyFile{tmpfn, runner.HostOutputDir + bind[len(runner.Container.OutputPath):]})
 610                         } else {
 611                                 // If a secret is outside OutputPath,
 612                                 // we bind mount the secret file
 613                                 // directly just like other mounts. We
 614                                 // also use this strategy when a
 615                                 // secret is inside OutputPath but
 616                                 // OutputPath is a live collection, to
 617                                 // avoid writing the secret to
 618                                 // Keep. Attempting to remove a
 619                                 // bind-mounted secret file from
 620                                 // inside the container will return a
 621                                 // "Device or resource busy" error
 622                                 // that might not be handled well by
 623                                 // the container, which is why we
 624                                 // don't use this strategy when
 625                                 // OutputPath is a staging directory.
 626                                 bindmounts[bind] = bindmount{HostPath: tmpfn, ReadOnly: true}
 627                         }
 628
 629                 case mnt.Kind == "git_tree":
 630                         tmpdir, err := runner.MkTempDir(runner.parentTemp, "git_tree")
 631                         if err != nil {
 632                                 return nil, fmt.Errorf("creating temp dir: %v", err)
 633                         }
 634                         err = gitMount(mnt).extractTree(runner.containerClient, tmpdir, token)
 635                         if err != nil {
 636                                 return nil, err
 637                         }
 638                         bindmounts[bind] = bindmount{HostPath: tmpdir, ReadOnly: true}
 639                 }
 640         }
 641
 642         if runner.HostOutputDir == "" {
 643                 return nil, fmt.Errorf("output path does not correspond to a writable mount point")
 644         }
 645
 646         if needCertMount && runner.Container.RuntimeConstraints.API {
 647                 for _, certfile := range arvadosclient.CertFiles {
 648                         _, err := os.Stat(certfile)
 649                         if err == nil {
 650                                 bindmounts["/etc/arvados/ca-certificates.crt"] = bindmount{HostPath: certfile, ReadOnly: true}
 651                                 break
 652                         }
 653                 }
 654         }
 655
 656         if pdhOnly {
 657                 // If we are only mounting collections by pdh, make
 658                 // sure we don't subscribe to websocket events to
 659                 // avoid putting undesired load on the API server
 660                 arvMountCmd = append(arvMountCmd, "--mount-by-pdh", "by_id", "--disable-event-listening")
 661         } else {
 662                 arvMountCmd = append(arvMountCmd, "--mount-by-id", "by_id")
 663         }
 664         // the by_uuid mount point is used by singularity when writing
 665         // out docker images converted to SIF
 666         arvMountCmd = append(arvMountCmd, "--mount-by-id", "by_uuid")
 667         arvMountCmd = append(arvMountCmd, runner.ArvMountPoint)
 668
 669         runner.ArvMount, err = runner.RunArvMount(arvMountCmd, token)
 670         if err != nil {
 671                 return nil, fmt.Errorf("while trying to start arv-mount: %v", err)
 672         }
 673         if runner.hoststatReporter != nil && runner.ArvMount != nil {
 674                 runner.hoststatReporter.ReportPID("arv-mount", runner.ArvMount.Process.Pid)
 675         }
 676
 677         for _, p := range collectionPaths {
 678                 _, err = os.Stat(p)
 679                 if err != nil {
 680                         return nil, fmt.Errorf("while checking that input files exist: %v", err)
 681                 }
 682         }
 683
 684         for _, cp := range copyFiles {
 685                 st, err := os.Stat(cp.src)
 686                 if err != nil {
 687                         return nil, fmt.Errorf("while staging writable file from %q to %q: %v", cp.src, cp.bind, err)
 688                 }
 689                 if st.IsDir() {
 690                         err = filepath.Walk(cp.src, func(walkpath string, walkinfo os.FileInfo, walkerr error) error {
 691                                 if walkerr != nil {
 692                                         return walkerr
 693                                 }
 694                                 target := path.Join(cp.bind, walkpath[len(cp.src):])
 695                                 if walkinfo.Mode().IsRegular() {
 696                                         copyerr := copyfile(walkpath, target)
 697                                         if copyerr != nil {
 698                                                 return copyerr
 699                                         }
 700                                         return os.Chmod(target, walkinfo.Mode()|0777)
 701                                 } else if walkinfo.Mode().IsDir() {
 702                                         mkerr := os.MkdirAll(target, 0777)
 703                                         if mkerr != nil {
 704                                                 return mkerr
 705                                         }
 706                                         return os.Chmod(target, walkinfo.Mode()|os.ModeSetgid|0777)
 707                                 } else {
 708                                         return fmt.Errorf("source %q is not a regular file or directory", cp.src)
 709                                 }
 710                         })
 711                 } else if st.Mode().IsRegular() {
 712                         err = copyfile(cp.src, cp.bind)
 713                         if err == nil {
 714                                 err = os.Chmod(cp.bind, st.Mode()|0777)
 715                         }
 716                 }
 717                 if err != nil {
 718                         return nil, fmt.Errorf("while staging writable file from %q to %q: %v", cp.src, cp.bind, err)
 719                 }
 720         }
 721
 722         return bindmounts, nil
 723 }
 724
 725 func (runner *ContainerRunner) stopHoststat() error {
 726         if runner.hoststatReporter == nil {
 727                 return nil
 728         }
 729         runner.hoststatReporter.Stop()
 730         runner.hoststatReporter.LogProcessMemMax(runner.CrunchLog)
 731         err := runner.hoststatLogger.Close()
 732         if err != nil {
 733                 return fmt.Errorf("error closing hoststat logs: %v", err)
 734         }
 735         return nil
 736 }
 737
 738 func (runner *ContainerRunner) startHoststat() error {
 739         w, err := runner.NewLogWriter("hoststat")
 740         if err != nil {
 741                 return err
 742         }
 743         runner.hoststatLogger = NewThrottledLogger(w)
 744         runner.hoststatReporter = &crunchstat.Reporter{
 745                 Logger: log.New(runner.hoststatLogger, "", 0),
 746                 // Our own cgroup is the "host" cgroup, in the sense
 747                 // that it accounts for resource usage outside the
 748                 // container. It doesn't count _all_ resource usage on
 749                 // the system.
 750                 //
 751                 // TODO?: Use the furthest ancestor of our own cgroup
 752                 // that has stats available. (Currently crunchstat
 753                 // does not have that capability.)
 754                 Pid:        os.Getpid,
 755                 PollPeriod: runner.statInterval,
 756         }
 757         runner.hoststatReporter.Start()
 758         runner.hoststatReporter.ReportPID("crunch-run", os.Getpid())
 759         return nil
 760 }
 761
 762 func (runner *ContainerRunner) startCrunchstat() error {
 763         w, err := runner.NewLogWriter("crunchstat")
 764         if err != nil {
 765                 return err
 766         }
 767         runner.statLogger = NewThrottledLogger(w)
 768         runner.statReporter = &crunchstat.Reporter{
 769                 Pid:    runner.executor.Pid,
 770                 FS:     runner.crunchstatFakeFS,
 771                 Logger: log.New(runner.statLogger, "", 0),
 772                 MemThresholds: map[string][]crunchstat.Threshold{
 773                         "rss": crunchstat.NewThresholdsFromPercentages(runner.Container.RuntimeConstraints.RAM, []int64{90, 95, 99}),
 774                 },
 775                 PollPeriod:      runner.statInterval,
 776                 TempDir:         runner.parentTemp,
 777                 ThresholdLogger: runner.CrunchLog,
 778         }
 779         runner.statReporter.Start()
 780         return nil
 781 }
 782
 783 type infoCommand struct {
 784         label string
 785         cmd   []string
 786 }
 787
 788 // LogHostInfo logs info about the current host, for debugging and
 789 // accounting purposes. Although it's logged as "node-info", this is
 790 // about the environment where crunch-run is actually running, which
 791 // might differ from what's described in the node record (see
 792 // LogNodeRecord).
 793 func (runner *ContainerRunner) LogHostInfo() (err error) {
 794         w, err := runner.NewLogWriter("node-info")
 795         if err != nil {
 796                 return
 797         }
 798
 799         commands := []infoCommand{
 800                 {
 801                         label: "Host Information",
 802                         cmd:   []string{"uname", "-a"},
 803                 },
 804                 {
 805                         label: "CPU Information",
 806                         cmd:   []string{"cat", "/proc/cpuinfo"},
 807                 },
 808                 {
 809                         label: "Memory Information",
 810                         cmd:   []string{"cat", "/proc/meminfo"},
 811                 },
 812                 {
 813                         label: "Disk Space",
 814                         cmd:   []string{"df", "-m", "/", os.TempDir()},
 815                 },
 816                 {
 817                         label: "Disk INodes",
 818                         cmd:   []string{"df", "-i", "/", os.TempDir()},
 819                 },
 820         }
 821
 822         // Run commands with informational output to be logged.
 823         for _, command := range commands {
 824                 fmt.Fprintln(w, command.label)
 825                 cmd := exec.Command(command.cmd[0], command.cmd[1:]...)
 826                 cmd.Stdout = w
 827                 cmd.Stderr = w
 828                 if err := cmd.Run(); err != nil {
 829                         err = fmt.Errorf("While running command %q: %v", command.cmd, err)
 830                         fmt.Fprintln(w, err)
 831                         return err
 832                 }
 833                 fmt.Fprintln(w, "")
 834         }
 835
 836         err = w.Close()
 837         if err != nil {
 838                 return fmt.Errorf("While closing node-info logs: %v", err)
 839         }
 840         return nil
 841 }
 842
 843 // LogContainerRecord gets and saves the raw JSON container record from the API server
 844 func (runner *ContainerRunner) LogContainerRecord() error {
 845         logged, err := runner.logAPIResponse("container", "containers", map[string]interface{}{"filters": [][]string{{"uuid", "=", runner.Container.UUID}}}, nil)
 846         if !logged && err == nil {
 847                 err = fmt.Errorf("error: no container record found for %s", runner.Container.UUID)
 848         }
 849         return err
 850 }
 851
 852 // LogNodeRecord logs the current host's InstanceType config entry (or
 853 // the arvados#node record, if running via crunch-dispatch-slurm).
 854 func (runner *ContainerRunner) LogNodeRecord() error {
 855         if it := os.Getenv("InstanceType"); it != "" {
 856                 // Dispatched via arvados-dispatch-cloud. Save
 857                 // InstanceType config fragment received from
 858                 // dispatcher on stdin.
 859                 w, err := runner.LogCollection.OpenFile("node.json", os.O_CREATE|os.O_WRONLY, 0666)
 860                 if err != nil {
 861                         return err
 862                 }
 863                 defer w.Close()
 864                 _, err = io.WriteString(w, it)
 865                 if err != nil {
 866                         return err
 867                 }
 868                 return w.Close()
 869         }
 870         // Dispatched via crunch-dispatch-slurm. Look up
 871         // apiserver's node record corresponding to
 872         // $SLURMD_NODENAME.
 873         hostname := os.Getenv("SLURMD_NODENAME")
 874         if hostname == "" {
 875                 hostname, _ = os.Hostname()
 876         }
 877         _, err := runner.logAPIResponse("node", "nodes", map[string]interface{}{"filters": [][]string{{"hostname", "=", hostname}}}, func(resp interface{}) {
 878                 // The "info" field has admin-only info when
 879                 // obtained with a privileged token, and
 880                 // should not be logged.
 881                 node, ok := resp.(map[string]interface{})
 882                 if ok {
 883                         delete(node, "info")
 884                 }
 885         })
 886         return err
 887 }
 888
 889 func (runner *ContainerRunner) logAPIResponse(label, path string, params map[string]interface{}, munge func(interface{})) (logged bool, err error) {
 890         writer, err := runner.LogCollection.OpenFile(label+".json", os.O_CREATE|os.O_WRONLY, 0666)
 891         if err != nil {
 892                 return false, err
 893         }
 894         w := &ArvLogWriter{
 895                 ArvClient:     runner.DispatcherArvClient,
 896                 UUID:          runner.Container.UUID,
 897                 loggingStream: label,
 898                 writeCloser:   writer,
 899         }
 900
 901         reader, err := runner.DispatcherArvClient.CallRaw("GET", path, "", "", arvadosclient.Dict(params))
 902         if err != nil {
 903                 return false, fmt.Errorf("error getting %s record: %v", label, err)
 904         }
 905         defer reader.Close()
 906
 907         dec := json.NewDecoder(reader)
 908         dec.UseNumber()
 909         var resp map[string]interface{}
 910         if err = dec.Decode(&resp); err != nil {
 911                 return false, fmt.Errorf("error decoding %s list response: %v", label, err)
 912         }
 913         items, ok := resp["items"].([]interface{})
 914         if !ok {
 915                 return false, fmt.Errorf("error decoding %s list response: no \"items\" key in API list response", label)
 916         } else if len(items) < 1 {
 917                 return false, nil
 918         }
 919         if munge != nil {
 920                 munge(items[0])
 921         }
 922         // Re-encode it using indentation to improve readability
 923         enc := json.NewEncoder(w)
 924         enc.SetIndent("", "    ")
 925         if err = enc.Encode(items[0]); err != nil {
 926                 return false, fmt.Errorf("error logging %s record: %v", label, err)
 927         }
 928         err = w.Close()
 929         if err != nil {
 930                 return false, fmt.Errorf("error closing %s.json in log collection: %v", label, err)
 931         }
 932         return true, nil
 933 }
 934
 935 func (runner *ContainerRunner) getStdoutFile(mntPath string) (*os.File, error) {
 936         stdoutPath := mntPath[len(runner.Container.OutputPath):]
 937         index := strings.LastIndex(stdoutPath, "/")
 938         if index > 0 {
 939                 subdirs := stdoutPath[:index]
 940                 if subdirs != "" {
 941                         st, err := os.Stat(runner.HostOutputDir)
 942                         if err != nil {
 943                                 return nil, fmt.Errorf("While Stat on temp dir: %v", err)
 944                         }
 945                         stdoutPath := filepath.Join(runner.HostOutputDir, subdirs)
 946                         err = os.MkdirAll(stdoutPath, st.Mode()|os.ModeSetgid|0777)
 947                         if err != nil {
 948                                 return nil, fmt.Errorf("While MkdirAll %q: %v", stdoutPath, err)
 949                         }
 950                 }
 951         }
 952         stdoutFile, err := os.Create(filepath.Join(runner.HostOutputDir, stdoutPath))
 953         if err != nil {
 954                 return nil, fmt.Errorf("While creating file %q: %v", stdoutPath, err)
 955         }
 956
 957         return stdoutFile, nil
 958 }
 959
 960 // CreateContainer creates the docker container.
 961 func (runner *ContainerRunner) CreateContainer(imageID string, bindmounts map[string]bindmount) error {
 962         var stdin io.ReadCloser = ioutil.NopCloser(bytes.NewReader(nil))
 963         if mnt, ok := runner.Container.Mounts["stdin"]; ok {
 964                 switch mnt.Kind {
 965                 case "collection":
 966                         var collID string
 967                         if mnt.UUID != "" {
 968                                 collID = mnt.UUID
 969                         } else {
 970                                 collID = mnt.PortableDataHash
 971                         }
 972                         path := runner.ArvMountPoint + "/by_id/" + collID + "/" + mnt.Path
 973                         f, err := os.Open(path)
 974                         if err != nil {
 975                                 return err
 976                         }
 977                         stdin = f
 978                 case "json":
 979                         j, err := json.Marshal(mnt.Content)
 980                         if err != nil {
 981                                 return fmt.Errorf("error encoding stdin json data: %v", err)
 982                         }
 983                         stdin = ioutil.NopCloser(bytes.NewReader(j))
 984                 default:
 985                         return fmt.Errorf("stdin mount has unsupported kind %q", mnt.Kind)
 986                 }
 987         }
 988
 989         var stdout, stderr io.WriteCloser
 990         if mnt, ok := runner.Container.Mounts["stdout"]; ok {
 991                 f, err := runner.getStdoutFile(mnt.Path)
 992                 if err != nil {
 993                         return err
 994                 }
 995                 stdout = f
 996         } else if w, err := runner.NewLogWriter("stdout"); err != nil {
 997                 return err
 998         } else {
 999                 stdout = NewThrottledLogger(w)
1000         }
1001
1002         if mnt, ok := runner.Container.Mounts["stderr"]; ok {
1003                 f, err := runner.getStdoutFile(mnt.Path)
1004                 if err != nil {
1005                         return err
1006                 }
1007                 stderr = f
1008         } else if w, err := runner.NewLogWriter("stderr"); err != nil {
1009                 return err
1010         } else {
1011                 stderr = NewThrottledLogger(w)
1012         }
1013
1014         env := runner.Container.Environment
1015         enableNetwork := runner.enableNetwork == "always"
1016         if runner.Container.RuntimeConstraints.API {
1017                 enableNetwork = true
1018                 tok, err := runner.ContainerToken()
1019                 if err != nil {
1020                         return err
1021                 }
1022                 env = map[string]string{}
1023                 for k, v := range runner.Container.Environment {
1024                         env[k] = v
1025                 }
1026                 env["ARVADOS_API_TOKEN"] = tok
1027                 env["ARVADOS_API_HOST"] = os.Getenv("ARVADOS_API_HOST")
1028                 env["ARVADOS_API_HOST_INSECURE"] = os.Getenv("ARVADOS_API_HOST_INSECURE")
1029                 env["ARVADOS_KEEP_SERVICES"] = os.Getenv("ARVADOS_KEEP_SERVICES")
1030         }
1031         workdir := runner.Container.Cwd
1032         if workdir == "." {
1033                 // both "" and "." mean default
1034                 workdir = ""
1035         }
1036         ram := runner.Container.RuntimeConstraints.RAM
1037         if !runner.enableMemoryLimit {
1038                 ram = 0
1039         }
1040         runner.executorStdin = stdin
1041         runner.executorStdout = stdout
1042         runner.executorStderr = stderr
1043
1044         if runner.Container.RuntimeConstraints.CUDA.DeviceCount > 0 {
1045                 nvidiaModprobe(runner.CrunchLog)
1046         }
1047
1048         return runner.executor.Create(containerSpec{
1049                 Image:           imageID,
1050                 VCPUs:           runner.Container.RuntimeConstraints.VCPUs,
1051                 RAM:             ram,
1052                 WorkingDir:      workdir,
1053                 Env:             env,
1054                 BindMounts:      bindmounts,
1055                 Command:         runner.Container.Command,
1056                 EnableNetwork:   enableNetwork,
1057                 CUDADeviceCount: runner.Container.RuntimeConstraints.CUDA.DeviceCount,
1058                 NetworkMode:     runner.networkMode,
1059                 CgroupParent:    runner.setCgroupParent,
1060                 Stdin:           stdin,
1061                 Stdout:          stdout,
1062                 Stderr:          stderr,
1063         })
1064 }
1065
1066 // StartContainer starts the docker container created by CreateContainer.
1067 func (runner *ContainerRunner) StartContainer() error {
1068         runner.CrunchLog.Printf("Starting container")
1069         runner.cStateLock.Lock()
1070         defer runner.cStateLock.Unlock()
1071         if runner.cCancelled {
1072                 return ErrCancelled
1073         }
1074         err := runner.executor.Start()
1075         if err != nil {
1076                 var advice string
1077                 if m, e := regexp.MatchString("(?ms).*(exec|System error).*(no such file or directory|file not found).*", err.Error()); m && e == nil {
1078                         advice = fmt.Sprintf("\nPossible causes: command %q is missing, the interpreter given in #! is missing, or script has Windows line endings.", runner.Container.Command[0])
1079                 }
1080                 return fmt.Errorf("could not start container: %v%s", err, advice)
1081         }
1082         return nil
1083 }
1084
1085 // WaitFinish waits for the container to terminate, capture the exit code, and
1086 // close the stdout/stderr logging.
1087 func (runner *ContainerRunner) WaitFinish() error {
1088         runner.CrunchLog.Print("Waiting for container to finish")
1089         var timeout <-chan time.Time
1090         if s := runner.Container.SchedulingParameters.MaxRunTime; s > 0 {
1091                 timeout = time.After(time.Duration(s) * time.Second)
1092         }
1093         ctx, cancel := context.WithCancel(context.Background())
1094         defer cancel()
1095         go func() {
1096                 select {
1097                 case <-timeout:
1098                         runner.CrunchLog.Printf("maximum run time exceeded. Stopping container.")
1099                         runner.stop(nil)
1100                 case <-runner.ArvMountExit:
1101                         runner.CrunchLog.Printf("arv-mount exited while container is still running. Stopping container.")
1102                         runner.stop(nil)
1103                 case <-ctx.Done():
1104                 }
1105         }()
1106         exitcode, err := runner.executor.Wait(ctx)
1107         if err != nil {
1108                 runner.checkBrokenNode(err)
1109                 return err
1110         }
1111         runner.ExitCode = &exitcode
1112
1113         extra := ""
1114         if exitcode&0x80 != 0 {
1115                 // Convert raw exit status (0x80 + signal number) to a
1116                 // string to log after the code, like " (signal 101)"
1117                 // or " (signal 9, killed)"
1118                 sig := syscall.WaitStatus(exitcode).Signal()
1119                 if name := unix.SignalName(sig); name != "" {
1120                         extra = fmt.Sprintf(" (signal %d, %s)", sig, name)
1121                 } else {
1122                         extra = fmt.Sprintf(" (signal %d)", sig)
1123                 }
1124         }
1125         runner.CrunchLog.Printf("Container exited with status code %d%s", exitcode, extra)
1126         err = runner.DispatcherArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{
1127                 "select":    []string{"uuid"},
1128                 "container": arvadosclient.Dict{"exit_code": exitcode},
1129         }, nil)
1130         if err != nil {
1131                 runner.CrunchLog.Printf("ignoring error updating exit_code: %s", err)
1132         }
1133
1134         var returnErr error
1135         if err = runner.executorStdin.Close(); err != nil {
1136                 err = fmt.Errorf("error closing container stdin: %s", err)
1137                 runner.CrunchLog.Printf("%s", err)
1138                 returnErr = err
1139         }
1140         if err = runner.executorStdout.Close(); err != nil {
1141                 err = fmt.Errorf("error closing container stdout: %s", err)
1142                 runner.CrunchLog.Printf("%s", err)
1143                 if returnErr == nil {
1144                         returnErr = err
1145                 }
1146         }
1147         if err = runner.executorStderr.Close(); err != nil {
1148                 err = fmt.Errorf("error closing container stderr: %s", err)
1149                 runner.CrunchLog.Printf("%s", err)
1150                 if returnErr == nil {
1151                         returnErr = err
1152                 }
1153         }
1154
1155         if runner.statReporter != nil {
1156                 runner.statReporter.Stop()
1157                 runner.statReporter.LogMaxima(runner.CrunchLog, map[string]int64{
1158                         "rss": runner.Container.RuntimeConstraints.RAM,
1159                 })
1160                 err = runner.statLogger.Close()
1161                 if err != nil {
1162                         runner.CrunchLog.Printf("error closing crunchstat logs: %v", err)
1163                 }
1164         }
1165         return returnErr
1166 }
1167
1168 func (runner *ContainerRunner) updateLogs() {
1169         ticker := time.NewTicker(crunchLogUpdatePeriod / 360)
1170         defer ticker.Stop()
1171
1172         sigusr1 := make(chan os.Signal, 1)
1173         signal.Notify(sigusr1, syscall.SIGUSR1)
1174         defer signal.Stop(sigusr1)
1175
1176         saveAtTime := time.Now().Add(crunchLogUpdatePeriod)
1177         saveAtSize := crunchLogUpdateSize
1178         var savedSize int64
1179         for {
1180                 select {
1181                 case <-ticker.C:
1182                 case <-sigusr1:
1183                         saveAtTime = time.Now()
1184                 }
1185                 runner.logMtx.Lock()
1186                 done := runner.LogsPDH != nil
1187                 runner.logMtx.Unlock()
1188                 if done {
1189                         return
1190                 }
1191                 size := runner.LogCollection.Size()
1192                 if size == savedSize || (time.Now().Before(saveAtTime) && size < saveAtSize) {
1193                         continue
1194                 }
1195                 saveAtTime = time.Now().Add(crunchLogUpdatePeriod)
1196                 saveAtSize = runner.LogCollection.Size() + crunchLogUpdateSize
1197                 saved, err := runner.saveLogCollection(false)
1198                 if err != nil {
1199                         runner.CrunchLog.Printf("error updating log collection: %s", err)
1200                         continue
1201                 }
1202
1203                 err = runner.DispatcherArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{
1204                         "select": []string{"uuid"},
1205                         "container": arvadosclient.Dict{
1206                                 "log": saved.PortableDataHash,
1207                         },
1208                 }, nil)
1209                 if err != nil {
1210                         runner.CrunchLog.Printf("error updating container log to %s: %s", saved.PortableDataHash, err)
1211                         continue
1212                 }
1213
1214                 savedSize = size
1215         }
1216 }
1217
1218 var spotInterruptionCheckInterval = 5 * time.Second
1219 var ec2MetadataBaseURL = "http://169.254.169.254"
1220
1221 const ec2TokenTTL = time.Second * 21600
1222
1223 func (runner *ContainerRunner) checkSpotInterruptionNotices() {
1224         type ec2metadata struct {
1225                 Action string    `json:"action"`
1226                 Time   time.Time `json:"time"`
1227         }
1228         runner.CrunchLog.Printf("Checking for spot interruptions every %v using instance metadata at %s", spotInterruptionCheckInterval, ec2MetadataBaseURL)
1229         var metadata ec2metadata
1230         var token string
1231         var tokenExp time.Time
1232         check := func() error {
1233                 ctx, cancel := context.WithDeadline(context.Background(), time.Now().Add(time.Minute))
1234                 defer cancel()
1235                 if token == "" || tokenExp.Sub(time.Now()) < time.Minute {
1236                         req, err := http.NewRequestWithContext(ctx, http.MethodPut, ec2MetadataBaseURL+"/latest/api/token", nil)
1237                         if err != nil {
1238                                 return err
1239                         }
1240                         req.Header.Set("X-aws-ec2-metadata-token-ttl-seconds", fmt.Sprintf("%d", int(ec2TokenTTL/time.Second)))
1241                         resp, err := http.DefaultClient.Do(req)
1242                         if err != nil {
1243                                 return err
1244                         }
1245                         defer resp.Body.Close()
1246                         if resp.StatusCode != http.StatusOK {
1247                                 return fmt.Errorf("%s", resp.Status)
1248                         }
1249                         newtoken, err := ioutil.ReadAll(resp.Body)
1250                         if err != nil {
1251                                 return err
1252                         }
1253                         token = strings.TrimSpace(string(newtoken))
1254                         tokenExp = time.Now().Add(ec2TokenTTL)
1255                 }
1256                 req, err := http.NewRequestWithContext(ctx, http.MethodGet, ec2MetadataBaseURL+"/latest/meta-data/spot/instance-action", nil)
1257                 if err != nil {
1258                         return err
1259                 }
1260                 req.Header.Set("X-aws-ec2-metadata-token", token)
1261                 resp, err := http.DefaultClient.Do(req)
1262                 if err != nil {
1263                         return err
1264                 }
1265                 defer resp.Body.Close()
1266                 metadata = ec2metadata{}
1267                 switch resp.StatusCode {
1268                 case http.StatusOK:
1269                         break
1270                 case http.StatusNotFound:
1271                         // "If Amazon EC2 is not preparing to stop or
1272                         // terminate the instance, or if you
1273                         // terminated the instance yourself,
1274                         // instance-action is not present in the
1275                         // instance metadata and you receive an HTTP
1276                         // 404 error when you try to retrieve it."
1277                         return nil
1278                 case http.StatusUnauthorized:
1279                         token = ""
1280                         return fmt.Errorf("%s", resp.Status)
1281                 default:
1282                         return fmt.Errorf("%s", resp.Status)
1283                 }
1284                 err = json.NewDecoder(resp.Body).Decode(&metadata)
1285                 if err != nil {
1286                         return err
1287                 }
1288                 return nil
1289         }
1290         failures := 0
1291         var lastmetadata ec2metadata
1292         for range time.NewTicker(spotInterruptionCheckInterval).C {
1293                 err := check()
1294                 if err != nil {
1295                         runner.CrunchLog.Printf("Error checking spot interruptions: %s", err)
1296                         failures++
1297                         if failures > 5 {
1298                                 runner.CrunchLog.Printf("Giving up on checking spot interruptions after too many consecutive failures")
1299                                 return
1300                         }
1301                         continue
1302                 }
1303                 failures = 0
1304                 if metadata != lastmetadata {
1305                         lastmetadata = metadata
1306                         text := fmt.Sprintf("Cloud provider scheduled instance %s at %s", metadata.Action, metadata.Time.UTC().Format(time.RFC3339))
1307                         runner.CrunchLog.Printf("%s", text)
1308                         runner.updateRuntimeStatus(arvadosclient.Dict{
1309                                 "warning":          "preemption notice",
1310                                 "warningDetail":    text,
1311                                 "preemptionNotice": text,
1312                         })
1313                         if proc, err := os.FindProcess(os.Getpid()); err == nil {
1314                                 // trigger updateLogs
1315                                 proc.Signal(syscall.SIGUSR1)
1316                         }
1317                 }
1318         }
1319 }
1320
1321 func (runner *ContainerRunner) updateRuntimeStatus(status arvadosclient.Dict) {
1322         err := runner.DispatcherArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{
1323                 "select": []string{"uuid"},
1324                 "container": arvadosclient.Dict{
1325                         "runtime_status": status,
1326                 },
1327         }, nil)
1328         if err != nil {
1329                 runner.CrunchLog.Printf("error updating container runtime_status: %s", err)
1330         }
1331 }
1332
1333 // CaptureOutput saves data from the container's output directory if
1334 // needed, and updates the container output accordingly.
1335 func (runner *ContainerRunner) CaptureOutput(bindmounts map[string]bindmount) error {
1336         if runner.Container.RuntimeConstraints.API {
1337                 // Output may have been set directly by the container, so
1338                 // refresh the container record to check.
1339                 err := runner.DispatcherArvClient.Get("containers", runner.Container.UUID,
1340                         arvadosclient.Dict{
1341                                 "select": []string{"output"},
1342                         }, &runner.Container)
1343                 if err != nil {
1344                         return err
1345                 }
1346                 if runner.Container.Output != "" {
1347                         // Container output is already set.
1348                         runner.OutputPDH = &runner.Container.Output
1349                         return nil
1350                 }
1351         }
1352
1353         txt, err := (&copier{
1354                 client:        runner.containerClient,
1355                 keepClient:    runner.ContainerKeepClient,
1356                 hostOutputDir: runner.HostOutputDir,
1357                 ctrOutputDir:  runner.Container.OutputPath,
1358                 bindmounts:    bindmounts,
1359                 mounts:        runner.Container.Mounts,
1360                 secretMounts:  runner.SecretMounts,
1361                 logger:        runner.CrunchLog,
1362         }).Copy()
1363         if err != nil {
1364                 return err
1365         }
1366         if n := len(regexp.MustCompile(` [0-9a-f]+\+\S*\+R`).FindAllStringIndex(txt, -1)); n > 0 {
1367                 runner.CrunchLog.Printf("Copying %d data blocks from remote input collections...", n)
1368                 fs, err := (&arvados.Collection{ManifestText: txt}).FileSystem(runner.containerClient, runner.ContainerKeepClient)
1369                 if err != nil {
1370                         return err
1371                 }
1372                 txt, err = fs.MarshalManifest(".")
1373                 if err != nil {
1374                         return err
1375                 }
1376         }
1377         var resp arvados.Collection
1378         err = runner.ContainerArvClient.Create("collections", arvadosclient.Dict{
1379                 "ensure_unique_name": true,
1380                 "select":             []string{"portable_data_hash"},
1381                 "collection": arvadosclient.Dict{
1382                         "is_trashed":    true,
1383                         "name":          "output for " + runner.Container.UUID,
1384                         "manifest_text": txt,
1385                 },
1386         }, &resp)
1387         if err != nil {
1388                 return fmt.Errorf("error creating output collection: %v", err)
1389         }
1390         runner.OutputPDH = &resp.PortableDataHash
1391         return nil
1392 }
1393
1394 func (runner *ContainerRunner) CleanupDirs() {
1395         if runner.ArvMount != nil {
1396                 var delay int64 = 8
1397                 umount := exec.Command("arv-mount", fmt.Sprintf("--unmount-timeout=%d", delay), "--unmount", runner.ArvMountPoint)
1398                 umount.Stdout = runner.CrunchLog
1399                 umount.Stderr = runner.CrunchLog
1400                 runner.CrunchLog.Printf("Running %v", umount.Args)
1401                 umnterr := umount.Start()
1402
1403                 if umnterr != nil {
1404                         runner.CrunchLog.Printf("Error unmounting: %v", umnterr)
1405                         runner.ArvMount.Process.Kill()
1406                 } else {
1407                         // If arv-mount --unmount gets stuck for any reason, we
1408                         // don't want to wait for it forever.  Do Wait() in a goroutine
1409                         // so it doesn't block crunch-run.
1410                         umountExit := make(chan error)
1411                         go func() {
1412                                 mnterr := umount.Wait()
1413                                 if mnterr != nil {
1414                                         runner.CrunchLog.Printf("Error unmounting: %v", mnterr)
1415                                 }
1416                                 umountExit <- mnterr
1417                         }()
1418
1419                         for again := true; again; {
1420                                 again = false
1421                                 select {
1422                                 case <-umountExit:
1423                                         umount = nil
1424                                         again = true
1425                                 case <-runner.ArvMountExit:
1426                                         break
1427                                 case <-time.After(time.Duration((delay + 1) * int64(time.Second))):
1428                                         runner.CrunchLog.Printf("Timed out waiting for unmount")
1429                                         if umount != nil {
1430                                                 umount.Process.Kill()
1431                                         }
1432                                         runner.ArvMount.Process.Kill()
1433                                 }
1434                         }
1435                 }
1436                 runner.ArvMount = nil
1437         }
1438
1439         if runner.ArvMountPoint != "" {
1440                 if rmerr := os.Remove(runner.ArvMountPoint); rmerr != nil {
1441                         runner.CrunchLog.Printf("While cleaning up arv-mount directory %s: %v", runner.ArvMountPoint, rmerr)
1442                 }
1443                 runner.ArvMountPoint = ""
1444         }
1445
1446         if rmerr := os.RemoveAll(runner.parentTemp); rmerr != nil {
1447                 runner.CrunchLog.Printf("While cleaning up temporary directory %s: %v", runner.parentTemp, rmerr)
1448         }
1449 }
1450
1451 // CommitLogs posts the collection containing the final container logs.
1452 func (runner *ContainerRunner) CommitLogs() error {
1453         func() {
1454                 // Hold cStateLock to prevent races on CrunchLog (e.g., stop()).
1455                 runner.cStateLock.Lock()
1456                 defer runner.cStateLock.Unlock()
1457
1458                 runner.CrunchLog.Print(runner.finalState)
1459
1460                 if runner.arvMountLog != nil {
1461                         runner.arvMountLog.Close()
1462                 }
1463                 runner.CrunchLog.Close()
1464
1465                 // Closing CrunchLog above allows them to be committed to Keep at this
1466                 // point, but re-open crunch log with ArvClient in case there are any
1467                 // other further errors (such as failing to write the log to Keep!)
1468                 // while shutting down
1469                 runner.CrunchLog = NewThrottledLogger(&ArvLogWriter{
1470                         ArvClient:     runner.DispatcherArvClient,
1471                         UUID:          runner.Container.UUID,
1472                         loggingStream: "crunch-run",
1473                         writeCloser:   nil,
1474                 })
1475                 runner.CrunchLog.Immediate = log.New(os.Stderr, runner.Container.UUID+" ", 0)
1476         }()
1477
1478         if runner.keepstoreLogger != nil {
1479                 // Flush any buffered logs from our local keepstore
1480                 // process.  Discard anything logged after this point
1481                 // -- it won't end up in the log collection, so
1482                 // there's no point writing it to the collectionfs.
1483                 runner.keepstoreLogbuf.SetWriter(io.Discard)
1484                 runner.keepstoreLogger.Close()
1485                 runner.keepstoreLogger = nil
1486         }
1487
1488         if runner.LogsPDH != nil {
1489                 // If we have already assigned something to LogsPDH,
1490                 // we must be closing the re-opened log, which won't
1491                 // end up getting attached to the container record and
1492                 // therefore doesn't need to be saved as a collection
1493                 // -- it exists only to send logs to other channels.
1494                 return nil
1495         }
1496
1497         saved, err := runner.saveLogCollection(true)
1498         if err != nil {
1499                 return fmt.Errorf("error saving log collection: %s", err)
1500         }
1501         runner.logMtx.Lock()
1502         defer runner.logMtx.Unlock()
1503         runner.LogsPDH = &saved.PortableDataHash
1504         return nil
1505 }
1506
1507 // Create/update the log collection. Return value has UUID and
1508 // PortableDataHash fields populated, but others may be blank.
1509 func (runner *ContainerRunner) saveLogCollection(final bool) (response arvados.Collection, err error) {
1510         runner.logMtx.Lock()
1511         defer runner.logMtx.Unlock()
1512         if runner.LogsPDH != nil {
1513                 // Already finalized.
1514                 return
1515         }
1516         updates := arvadosclient.Dict{
1517                 "name": "logs for " + runner.Container.UUID,
1518         }
1519         mt, err1 := runner.LogCollection.MarshalManifest(".")
1520         if err1 == nil {
1521                 // Only send updated manifest text if there was no
1522                 // error.
1523                 updates["manifest_text"] = mt
1524         }
1525
1526         // Even if flushing the manifest had an error, we still want
1527         // to update the log record, if possible, to push the trash_at
1528         // and delete_at times into the future.  Details on bug
1529         // #17293.
1530         if final {
1531                 updates["is_trashed"] = true
1532         } else {
1533                 // We set trash_at so this collection gets
1534                 // automatically cleaned up eventually.  It used to be
1535                 // 12 hours but we had a situation where the API
1536                 // server was down over a weekend but the containers
1537                 // kept running such that the log collection got
1538                 // trashed, so now we make it 2 weeks.  refs #20378
1539                 exp := time.Now().Add(time.Duration(24*14) * time.Hour)
1540                 updates["trash_at"] = exp
1541                 updates["delete_at"] = exp
1542         }
1543         reqBody := arvadosclient.Dict{
1544                 "select":     []string{"uuid", "portable_data_hash"},
1545                 "collection": updates,
1546         }
1547         var err2 error
1548         if runner.logUUID == "" {
1549                 reqBody["ensure_unique_name"] = true
1550                 err2 = runner.DispatcherArvClient.Create("collections", reqBody, &response)
1551         } else {
1552                 err2 = runner.DispatcherArvClient.Update("collections", runner.logUUID, reqBody, &response)
1553         }
1554         if err2 == nil {
1555                 runner.logUUID = response.UUID
1556         }
1557
1558         if err1 != nil || err2 != nil {
1559                 err = fmt.Errorf("error recording logs: %q, %q", err1, err2)
1560         }
1561         return
1562 }
1563
1564 // UpdateContainerRunning updates the container state to "Running"
1565 func (runner *ContainerRunner) UpdateContainerRunning(logId string) error {
1566         runner.cStateLock.Lock()
1567         defer runner.cStateLock.Unlock()
1568         if runner.cCancelled {
1569                 return ErrCancelled
1570         }
1571         updates := arvadosclient.Dict{
1572                 "gateway_address": runner.gateway.Address,
1573                 "state":           "Running",
1574         }
1575         if logId != "" {
1576                 updates["log"] = logId
1577         }
1578         return runner.DispatcherArvClient.Update(
1579                 "containers",
1580                 runner.Container.UUID,
1581                 arvadosclient.Dict{
1582                         "select":    []string{"uuid"},
1583                         "container": updates,
1584                 },
1585                 nil,
1586         )
1587 }
1588
1589 // ContainerToken returns the api_token the container (and any
1590 // arv-mount processes) are allowed to use.
1591 func (runner *ContainerRunner) ContainerToken() (string, error) {
1592         if runner.token != "" {
1593                 return runner.token, nil
1594         }
1595
1596         var auth arvados.APIClientAuthorization
1597         err := runner.DispatcherArvClient.Call("GET", "containers", runner.Container.UUID, "auth", nil, &auth)
1598         if err != nil {
1599                 return "", err
1600         }
1601         runner.token = fmt.Sprintf("v2/%s/%s/%s", auth.UUID, auth.APIToken, runner.Container.UUID)
1602         return runner.token, nil
1603 }
1604
1605 // UpdateContainerFinal updates the container record state on API
1606 // server to "Complete" or "Cancelled"
1607 func (runner *ContainerRunner) UpdateContainerFinal() error {
1608         update := arvadosclient.Dict{}
1609         update["state"] = runner.finalState
1610         if runner.LogsPDH != nil {
1611                 update["log"] = *runner.LogsPDH
1612         }
1613         if runner.ExitCode != nil {
1614                 update["exit_code"] = *runner.ExitCode
1615         } else {
1616                 update["exit_code"] = nil
1617         }
1618         if runner.finalState == "Complete" && runner.OutputPDH != nil {
1619                 update["output"] = *runner.OutputPDH
1620         }
1621         update["cost"] = runner.calculateCost(time.Now())
1622         return runner.DispatcherArvClient.Update("containers", runner.Container.UUID, arvadosclient.Dict{
1623                 "select":    []string{"uuid"},
1624                 "container": update,
1625         }, nil)
1626 }
1627
1628 // IsCancelled returns the value of Cancelled, with goroutine safety.
1629 func (runner *ContainerRunner) IsCancelled() bool {
1630         runner.cStateLock.Lock()
1631         defer runner.cStateLock.Unlock()
1632         return runner.cCancelled
1633 }
1634
1635 // NewArvLogWriter creates an ArvLogWriter
1636 func (runner *ContainerRunner) NewArvLogWriter(name string) (io.WriteCloser, error) {
1637         writer, err := runner.LogCollection.OpenFile(name+".txt", os.O_CREATE|os.O_WRONLY, 0666)
1638         if err != nil {
1639                 return nil, err
1640         }
1641         return &ArvLogWriter{
1642                 ArvClient:     runner.DispatcherArvClient,
1643                 UUID:          runner.Container.UUID,
1644                 loggingStream: name,
1645                 writeCloser:   writer,
1646         }, nil
1647 }
1648
1649 // Run the full container lifecycle.
1650 func (runner *ContainerRunner) Run() (err error) {
1651         runner.CrunchLog.Printf("crunch-run %s started", cmd.Version.String())
1652         runner.CrunchLog.Printf("%s", currentUserAndGroups())
1653         v, _ := exec.Command("arv-mount", "--version").CombinedOutput()
1654         runner.CrunchLog.Printf("Using FUSE mount: %s", v)
1655         runner.CrunchLog.Printf("Using container runtime: %s", runner.executor.Runtime())
1656         runner.CrunchLog.Printf("Executing container: %s", runner.Container.UUID)
1657         runner.costStartTime = time.Now()
1658
1659         hostname, hosterr := os.Hostname()
1660         if hosterr != nil {
1661                 runner.CrunchLog.Printf("Error getting hostname '%v'", hosterr)
1662         } else {
1663                 runner.CrunchLog.Printf("Executing on host '%s'", hostname)
1664         }
1665
1666         sigusr2 := make(chan os.Signal, 1)
1667         signal.Notify(sigusr2, syscall.SIGUSR2)
1668         defer signal.Stop(sigusr2)
1669         runner.loadPrices()
1670         go runner.handleSIGUSR2(sigusr2)
1671
1672         runner.finalState = "Queued"
1673
1674         defer func() {
1675                 runner.CleanupDirs()
1676
1677                 runner.CrunchLog.Printf("crunch-run finished")
1678                 runner.CrunchLog.Close()
1679         }()
1680
1681         err = runner.fetchContainerRecord()
1682         if err != nil {
1683                 return
1684         }
1685         if runner.Container.State != "Locked" {
1686                 return fmt.Errorf("dispatch error detected: container %q has state %q", runner.Container.UUID, runner.Container.State)
1687         }
1688
1689         var bindmounts map[string]bindmount
1690         defer func() {
1691                 // checkErr prints e (unless it's nil) and sets err to
1692                 // e (unless err is already non-nil). Thus, if err
1693                 // hasn't already been assigned when Run() returns,
1694                 // this cleanup func will cause Run() to return the
1695                 // first non-nil error that is passed to checkErr().
1696                 checkErr := func(errorIn string, e error) {
1697                         if e == nil {
1698                                 return
1699                         }
1700                         runner.CrunchLog.Printf("error in %s: %v", errorIn, e)
1701                         if err == nil {
1702                                 err = e
1703                         }
1704                         if runner.finalState == "Complete" {
1705                                 // There was an error in the finalization.
1706                                 runner.finalState = "Cancelled"
1707                         }
1708                 }
1709
1710                 // Log the error encountered in Run(), if any
1711                 checkErr("Run", err)
1712
1713                 if runner.finalState == "Queued" {
1714                         runner.UpdateContainerFinal()
1715                         return
1716                 }
1717
1718                 if runner.IsCancelled() {
1719                         runner.finalState = "Cancelled"
1720                         // but don't return yet -- we still want to
1721                         // capture partial output and write logs
1722                 }
1723
1724                 if bindmounts != nil {
1725                         checkErr("CaptureOutput", runner.CaptureOutput(bindmounts))
1726                 }
1727                 checkErr("stopHoststat", runner.stopHoststat())
1728                 checkErr("CommitLogs", runner.CommitLogs())
1729                 runner.CleanupDirs()
1730                 checkErr("UpdateContainerFinal", runner.UpdateContainerFinal())
1731         }()
1732
1733         runner.setupSignals()
1734         err = runner.startHoststat()
1735         if err != nil {
1736                 return
1737         }
1738         if runner.keepstore != nil {
1739                 runner.hoststatReporter.ReportPID("keepstore", runner.keepstore.Process.Pid)
1740         }
1741
1742         // set up FUSE mount and binds
1743         bindmounts, err = runner.SetupMounts()
1744         if err != nil {
1745                 runner.finalState = "Cancelled"
1746                 err = fmt.Errorf("While setting up mounts: %v", err)
1747                 return
1748         }
1749
1750         // check for and/or load image
1751         imageID, err := runner.LoadImage()
1752         if err != nil {
1753                 if !runner.checkBrokenNode(err) {
1754                         // Failed to load image but not due to a "broken node"
1755                         // condition, probably user error.
1756                         runner.finalState = "Cancelled"
1757                 }
1758                 err = fmt.Errorf("While loading container image: %v", err)
1759                 return
1760         }
1761
1762         err = runner.CreateContainer(imageID, bindmounts)
1763         if err != nil {
1764                 return
1765         }
1766         err = runner.LogHostInfo()
1767         if err != nil {
1768                 return
1769         }
1770         err = runner.LogNodeRecord()
1771         if err != nil {
1772                 return
1773         }
1774         err = runner.LogContainerRecord()
1775         if err != nil {
1776                 return
1777         }
1778
1779         if runner.IsCancelled() {
1780                 return
1781         }
1782
1783         logCollection, err := runner.saveLogCollection(false)
1784         var logId string
1785         if err == nil {
1786                 logId = logCollection.PortableDataHash
1787         } else {
1788                 runner.CrunchLog.Printf("Error committing initial log collection: %v", err)
1789         }
1790         err = runner.UpdateContainerRunning(logId)
1791         if err != nil {
1792                 return
1793         }
1794         runner.finalState = "Cancelled"
1795
1796         err = runner.startCrunchstat()
1797         if err != nil {
1798                 return
1799         }
1800
1801         err = runner.StartContainer()
1802         if err != nil {
1803                 runner.checkBrokenNode(err)
1804                 return
1805         }
1806
1807         err = runner.WaitFinish()
1808         if err == nil && !runner.IsCancelled() {
1809                 runner.finalState = "Complete"
1810         }
1811         return
1812 }
1813
1814 // Fetch the current container record (uuid = runner.Container.UUID)
1815 // into runner.Container.
1816 func (runner *ContainerRunner) fetchContainerRecord() error {
1817         reader, err := runner.DispatcherArvClient.CallRaw("GET", "containers", runner.Container.UUID, "", nil)
1818         if err != nil {
1819                 return fmt.Errorf("error fetching container record: %v", err)
1820         }
1821         defer reader.Close()
1822
1823         dec := json.NewDecoder(reader)
1824         dec.UseNumber()
1825         err = dec.Decode(&runner.Container)
1826         if err != nil {
1827                 return fmt.Errorf("error decoding container record: %v", err)
1828         }
1829
1830         var sm struct {
1831                 SecretMounts map[string]arvados.Mount `json:"secret_mounts"`
1832         }
1833
1834         containerToken, err := runner.ContainerToken()
1835         if err != nil {
1836                 return fmt.Errorf("error getting container token: %v", err)
1837         }
1838
1839         runner.ContainerArvClient, runner.ContainerKeepClient,
1840                 runner.containerClient, err = runner.MkArvClient(containerToken)
1841         if err != nil {
1842                 return fmt.Errorf("error creating container API client: %v", err)
1843         }
1844
1845         runner.ContainerKeepClient.SetStorageClasses(runner.Container.OutputStorageClasses)
1846         runner.DispatcherKeepClient.SetStorageClasses(runner.Container.OutputStorageClasses)
1847
1848         err = runner.ContainerArvClient.Call("GET", "containers", runner.Container.UUID, "secret_mounts", nil, &sm)
1849         if err != nil {
1850                 if apierr, ok := err.(arvadosclient.APIServerError); !ok || apierr.HttpStatusCode != 404 {
1851                         return fmt.Errorf("error fetching secret_mounts: %v", err)
1852                 }
1853                 // ok && apierr.HttpStatusCode == 404, which means
1854                 // secret_mounts isn't supported by this API server.
1855         }
1856         runner.SecretMounts = sm.SecretMounts
1857
1858         return nil
1859 }
1860
1861 // NewContainerRunner creates a new container runner.
1862 func NewContainerRunner(dispatcherClient *arvados.Client,
1863         dispatcherArvClient IArvadosClient,
1864         dispatcherKeepClient IKeepClient,
1865         containerUUID string) (*ContainerRunner, error) {
1866
1867         cr := &ContainerRunner{
1868                 dispatcherClient:     dispatcherClient,
1869                 DispatcherArvClient:  dispatcherArvClient,
1870                 DispatcherKeepClient: dispatcherKeepClient,
1871         }
1872         cr.NewLogWriter = cr.NewArvLogWriter
1873         cr.RunArvMount = cr.ArvMountCmd
1874         cr.MkTempDir = ioutil.TempDir
1875         cr.MkArvClient = func(token string) (IArvadosClient, IKeepClient, *arvados.Client, error) {
1876                 cl, err := arvadosclient.MakeArvadosClient()
1877                 if err != nil {
1878                         return nil, nil, nil, err
1879                 }
1880                 cl.ApiToken = token
1881                 kc, err := keepclient.MakeKeepClient(cl)
1882                 if err != nil {
1883                         return nil, nil, nil, err
1884                 }
1885                 c2 := arvados.NewClientFromEnv()
1886                 c2.AuthToken = token
1887                 return cl, kc, c2, nil
1888         }
1889         var err error
1890         cr.LogCollection, err = (&arvados.Collection{}).FileSystem(cr.dispatcherClient, cr.DispatcherKeepClient)
1891         if err != nil {
1892                 return nil, err
1893         }
1894         cr.Container.UUID = containerUUID
1895         w, err := cr.NewLogWriter("crunch-run")
1896         if err != nil {
1897                 return nil, err
1898         }
1899         cr.CrunchLog = NewThrottledLogger(w)
1900         cr.CrunchLog.Immediate = log.New(os.Stderr, containerUUID+" ", 0)
1901
1902         loadLogThrottleParams(dispatcherArvClient)
1903         go cr.updateLogs()
1904
1905         return cr, nil
1906 }
1907
1908 func (command) RunCommand(prog string, args []string, stdin io.Reader, stdout, stderr io.Writer) int {
1909         log := log.New(stderr, "", 0)
1910         flags := flag.NewFlagSet(prog, flag.ContinueOnError)
1911         statInterval := flags.Duration("crunchstat-interval", 10*time.Second, "sampling period for periodic resource usage reporting")
1912         flags.String("cgroup-root", "/sys/fs/cgroup", "path to sysfs cgroup tree (obsolete, ignored)")
1913         flags.String("cgroup-parent", "docker", "name of container's parent cgroup (obsolete, ignored)")
1914         cgroupParentSubsystem := flags.String("cgroup-parent-subsystem", "", "use current cgroup for given `subsystem` as parent cgroup for container (subsystem argument is only relevant for cgroups v1; in cgroups v2 / unified mode, any non-empty value means use current cgroup); if empty, use the docker daemon's default cgroup parent. See https://doc.arvados.org/install/crunch2-slurm/install-dispatch.html#CrunchRunCommand-cgroups")
1915         caCertsPath := flags.String("ca-certs", "", "Path to TLS root certificates")
1916         detach := flags.Bool("detach", false, "Detach from parent process and run in the background")
1917         stdinConfig := flags.Bool("stdin-config", false, "Load config and environment variables from JSON message on stdin")
1918         configFile := flags.String("config", arvados.DefaultConfigFile, "filename of cluster config file to try loading if -stdin-config=false (default is $ARVADOS_CONFIG)")
1919         sleep := flags.Duration("sleep", 0, "Delay before starting (testing use only)")
1920         kill := flags.Int("kill", -1, "Send signal to an existing crunch-run process for given UUID")
1921         list := flags.Bool("list", false, "List UUIDs of existing crunch-run processes (and notify them to use price data passed on stdin)")
1922         enableMemoryLimit := flags.Bool("enable-memory-limit", true, "tell container runtime to limit container's memory usage")
1923         enableNetwork := flags.String("container-enable-networking", "default", "enable networking \"always\" (for all containers) or \"default\" (for containers that request it)")
1924         networkMode := flags.String("container-network-mode", "default", `Docker network mode for container (use any argument valid for docker --net)`)
1925         memprofile := flags.String("memprofile", "", "write memory profile to `file` after running container")
1926         runtimeEngine := flags.String("runtime-engine", "docker", "container runtime: docker or singularity")
1927         brokenNodeHook := flags.String("broken-node-hook", "", "script to run if node is detected to be broken (for example, Docker daemon is not running)")
1928         flags.Duration("check-containerd", 0, "Ignored. Exists for compatibility with older versions.")
1929         version := flags.Bool("version", false, "Write version information to stdout and exit 0.")
1930
1931         ignoreDetachFlag := false
1932         if len(args) > 0 && args[0] == "-no-detach" {
1933                 // This process was invoked by a parent process, which
1934                 // has passed along its own arguments, including
1935                 // -detach, after the leading -no-detach flag.  Strip
1936                 // the leading -no-detach flag (it's not recognized by
1937                 // flags.Parse()) and ignore the -detach flag that
1938                 // comes later.
1939                 args = args[1:]
1940                 ignoreDetachFlag = true
1941         }
1942
1943         if ok, code := cmd.ParseFlags(flags, prog, args, "container-uuid", stderr); !ok {
1944                 return code
1945         } else if *version {
1946                 fmt.Fprintln(stdout, prog, cmd.Version.String())
1947                 return 0
1948         } else if !*list && flags.NArg() != 1 {
1949                 fmt.Fprintf(stderr, "missing required argument: container-uuid (try -help)\n")
1950                 return 2
1951         }
1952
1953         containerUUID := flags.Arg(0)
1954
1955         switch {
1956         case *detach && !ignoreDetachFlag:
1957                 return Detach(containerUUID, prog, args, stdin, stdout, stderr)
1958         case *kill >= 0:
1959                 return KillProcess(containerUUID, syscall.Signal(*kill), stdout, stderr)
1960         case *list:
1961                 return ListProcesses(stdin, stdout, stderr)
1962         }
1963
1964         if len(containerUUID) != 27 {
1965                 log.Printf("usage: %s [options] UUID", prog)
1966                 return 1
1967         }
1968
1969         var keepstoreLogbuf bufThenWrite
1970         var conf ConfigData
1971         if *stdinConfig {
1972                 err := json.NewDecoder(stdin).Decode(&conf)
1973                 if err != nil {
1974                         log.Printf("decode stdin: %s", err)
1975                         return 1
1976                 }
1977                 for k, v := range conf.Env {
1978                         err = os.Setenv(k, v)
1979                         if err != nil {
1980                                 log.Printf("setenv(%q): %s", k, err)
1981                                 return 1
1982                         }
1983                 }
1984                 if conf.Cluster != nil {
1985                         // ClusterID is missing from the JSON
1986                         // representation, but we need it to generate
1987                         // a valid config file for keepstore, so we
1988                         // fill it using the container UUID prefix.
1989                         conf.Cluster.ClusterID = containerUUID[:5]
1990                 }
1991         } else {
1992                 conf = hpcConfData(containerUUID, *configFile, io.MultiWriter(&keepstoreLogbuf, stderr))
1993         }
1994
1995         log.Printf("crunch-run %s started", cmd.Version.String())
1996         time.Sleep(*sleep)
1997
1998         if *caCertsPath != "" {
1999                 arvadosclient.CertFiles = []string{*caCertsPath}
2000         }
2001
2002         keepstore, err := startLocalKeepstore(conf, io.MultiWriter(&keepstoreLogbuf, stderr))
2003         if err != nil {
2004                 log.Print(err)
2005                 return 1
2006         }
2007         if keepstore != nil {
2008                 defer keepstore.Process.Kill()
2009         }
2010
2011         api, err := arvadosclient.MakeArvadosClient()
2012         if err != nil {
2013                 log.Printf("%s: %v", containerUUID, err)
2014                 return 1
2015         }
2016         // arvadosclient now interprets Retries=10 to mean
2017         // Timeout=10m, retrying with exponential backoff + jitter.
2018         api.Retries = 10
2019
2020         kc, err := keepclient.MakeKeepClient(api)
2021         if err != nil {
2022                 log.Printf("%s: %v", containerUUID, err)
2023                 return 1
2024         }
2025         kc.BlockCache = &keepclient.BlockCache{MaxBlocks: 2}
2026         kc.Retries = 4
2027
2028         cr, err := NewContainerRunner(arvados.NewClientFromEnv(), api, kc, containerUUID)
2029         if err != nil {
2030                 log.Print(err)
2031                 return 1
2032         }
2033
2034         cr.keepstore = keepstore
2035         if keepstore == nil {
2036                 // Log explanation (if any) for why we're not running
2037                 // a local keepstore.
2038                 var buf bytes.Buffer
2039                 keepstoreLogbuf.SetWriter(&buf)
2040                 if buf.Len() > 0 {
2041                         cr.CrunchLog.Printf("%s", strings.TrimSpace(buf.String()))
2042                 }
2043         } else if logWhat := conf.Cluster.Containers.LocalKeepLogsToContainerLog; logWhat == "none" {
2044                 cr.CrunchLog.Printf("using local keepstore process (pid %d) at %s", keepstore.Process.Pid, os.Getenv("ARVADOS_KEEP_SERVICES"))
2045                 keepstoreLogbuf.SetWriter(io.Discard)
2046         } else {
2047                 cr.CrunchLog.Printf("using local keepstore process (pid %d) at %s, writing logs to keepstore.txt in log collection", keepstore.Process.Pid, os.Getenv("ARVADOS_KEEP_SERVICES"))
2048                 logwriter, err := cr.NewLogWriter("keepstore")
2049                 if err != nil {
2050                         log.Print(err)
2051                         return 1
2052                 }
2053                 cr.keepstoreLogger = NewThrottledLogger(logwriter)
2054
2055                 var writer io.WriteCloser = cr.keepstoreLogger
2056                 if logWhat == "errors" {
2057                         writer = &filterKeepstoreErrorsOnly{WriteCloser: writer}
2058                 } else if logWhat != "all" {
2059                         // should have been caught earlier by
2060                         // dispatcher's config loader
2061                         log.Printf("invalid value for Containers.LocalKeepLogsToContainerLog: %q", logWhat)
2062                         return 1
2063                 }
2064                 err = keepstoreLogbuf.SetWriter(writer)
2065                 if err != nil {
2066                         log.Print(err)
2067                         return 1
2068                 }
2069                 cr.keepstoreLogbuf = &keepstoreLogbuf
2070         }
2071
2072         switch *runtimeEngine {
2073         case "docker":
2074                 cr.executor, err = newDockerExecutor(containerUUID, cr.CrunchLog.Printf, cr.containerWatchdogInterval)
2075         case "singularity":
2076                 cr.executor, err = newSingularityExecutor(cr.CrunchLog.Printf)
2077         default:
2078                 cr.CrunchLog.Printf("%s: unsupported RuntimeEngine %q", containerUUID, *runtimeEngine)
2079                 cr.CrunchLog.Close()
2080                 return 1
2081         }
2082         if err != nil {
2083                 cr.CrunchLog.Printf("%s: %v", containerUUID, err)
2084                 cr.checkBrokenNode(err)
2085                 cr.CrunchLog.Close()
2086                 return 1
2087         }
2088         defer cr.executor.Close()
2089
2090         cr.brokenNodeHook = *brokenNodeHook
2091
2092         gwAuthSecret := os.Getenv("GatewayAuthSecret")
2093         os.Unsetenv("GatewayAuthSecret")
2094         if gwAuthSecret == "" {
2095                 // not safe to run a gateway service without an auth
2096                 // secret
2097                 cr.CrunchLog.Printf("Not starting a gateway server (GatewayAuthSecret was not provided by dispatcher)")
2098         } else {
2099                 gwListen := os.Getenv("GatewayAddress")
2100                 cr.gateway = Gateway{
2101                         Address:       gwListen,
2102                         AuthSecret:    gwAuthSecret,
2103                         ContainerUUID: containerUUID,
2104                         Target:        cr.executor,
2105                         Log:           cr.CrunchLog,
2106                         LogCollection: cr.LogCollection,
2107                 }
2108                 if gwListen == "" {
2109                         // Direct connection won't work, so we use the
2110                         // gateway_address field to indicate the
2111                         // internalURL of the controller process that
2112                         // has the current tunnel connection.
2113                         cr.gateway.ArvadosClient = cr.dispatcherClient
2114                         cr.gateway.UpdateTunnelURL = func(url string) {
2115                                 cr.gateway.Address = "tunnel " + url
2116                                 cr.DispatcherArvClient.Update("containers", containerUUID,
2117                                         arvadosclient.Dict{
2118                                                 "select":    []string{"uuid"},
2119                                                 "container": arvadosclient.Dict{"gateway_address": cr.gateway.Address},
2120                                         }, nil)
2121                         }
2122                 }
2123                 err = cr.gateway.Start()
2124                 if err != nil {
2125                         log.Printf("error starting gateway server: %s", err)
2126                         return 1
2127                 }
2128         }
2129
2130         parentTemp, tmperr := cr.MkTempDir("", "crunch-run."+containerUUID+".")
2131         if tmperr != nil {
2132                 log.Printf("%s: %v", containerUUID, tmperr)
2133                 return 1
2134         }
2135
2136         cr.parentTemp = parentTemp
2137         cr.statInterval = *statInterval
2138         cr.enableMemoryLimit = *enableMemoryLimit
2139         cr.enableNetwork = *enableNetwork
2140         cr.networkMode = *networkMode
2141         if *cgroupParentSubsystem != "" {
2142                 p, err := findCgroup(os.DirFS("/"), *cgroupParentSubsystem)
2143                 if err != nil {
2144                         log.Printf("fatal: cgroup parent subsystem: %s", err)
2145                         return 1
2146                 }
2147                 cr.setCgroupParent = p
2148         }
2149
2150         if conf.EC2SpotCheck {
2151                 go cr.checkSpotInterruptionNotices()
2152         }
2153
2154         runerr := cr.Run()
2155
2156         if *memprofile != "" {
2157                 f, err := os.Create(*memprofile)
2158                 if err != nil {
2159                         log.Printf("could not create memory profile: %s", err)
2160                 }
2161                 runtime.GC() // get up-to-date statistics
2162                 if err := pprof.WriteHeapProfile(f); err != nil {
2163                         log.Printf("could not write memory profile: %s", err)
2164                 }
2165                 closeerr := f.Close()
2166                 if closeerr != nil {
2167                         log.Printf("closing memprofile file: %s", err)
2168                 }
2169         }
2170
2171         if runerr != nil {
2172                 log.Printf("%s: %v", containerUUID, runerr)
2173                 return 1
2174         }
2175         return 0
2176 }
2177
2178 // Try to load ConfigData in hpc (slurm/lsf) environment. This means
2179 // loading the cluster config from the specified file and (if that
2180 // works) getting the runtime_constraints container field from
2181 // controller to determine # VCPUs so we can calculate KeepBuffers.
2182 func hpcConfData(uuid string, configFile string, stderr io.Writer) ConfigData {
2183         var conf ConfigData
2184         conf.Cluster = loadClusterConfigFile(configFile, stderr)
2185         if conf.Cluster == nil {
2186                 // skip loading the container record -- we won't be
2187                 // able to start local keepstore anyway.
2188                 return conf
2189         }
2190         arv, err := arvadosclient.MakeArvadosClient()
2191         if err != nil {
2192                 fmt.Fprintf(stderr, "error setting up arvadosclient: %s\n", err)
2193                 return conf
2194         }
2195         // arvadosclient now interprets Retries=10 to mean
2196         // Timeout=10m, retrying with exponential backoff + jitter.
2197         arv.Retries = 10
2198         var ctr arvados.Container
2199         err = arv.Call("GET", "containers", uuid, "", arvadosclient.Dict{"select": []string{"runtime_constraints"}}, &ctr)
2200         if err != nil {
2201                 fmt.Fprintf(stderr, "error getting container record: %s\n", err)
2202                 return conf
2203         }
2204         if ctr.RuntimeConstraints.VCPUs > 0 {
2205                 conf.KeepBuffers = ctr.RuntimeConstraints.VCPUs * conf.Cluster.Containers.LocalKeepBlobBuffersPerVCPU
2206         }
2207         return conf
2208 }
2209
2210 // Load cluster config file from given path. If an error occurs, log
2211 // the error to stderr and return nil.
2212 func loadClusterConfigFile(path string, stderr io.Writer) *arvados.Cluster {
2213         ldr := config.NewLoader(&bytes.Buffer{}, ctxlog.New(stderr, "plain", "info"))
2214         ldr.Path = path
2215         cfg, err := ldr.Load()
2216         if err != nil {
2217                 fmt.Fprintf(stderr, "could not load config file %s: %s\n", path, err)
2218                 return nil
2219         }
2220         cluster, err := cfg.GetCluster("")
2221         if err != nil {
2222                 fmt.Fprintf(stderr, "could not use config file %s: %s\n", path, err)
2223                 return nil
2224         }
2225         fmt.Fprintf(stderr, "loaded config file %s\n", path)
2226         return cluster
2227 }
2228
2229 func startLocalKeepstore(configData ConfigData, logbuf io.Writer) (*exec.Cmd, error) {
2230         if configData.KeepBuffers < 1 {
2231                 fmt.Fprintf(logbuf, "not starting a local keepstore process because KeepBuffers=%v in config\n", configData.KeepBuffers)
2232                 return nil, nil
2233         }
2234         if configData.Cluster == nil {
2235                 fmt.Fprint(logbuf, "not starting a local keepstore process because cluster config file was not loaded\n")
2236                 return nil, nil
2237         }
2238         for uuid, vol := range configData.Cluster.Volumes {
2239                 if len(vol.AccessViaHosts) > 0 {
2240                         fmt.Fprintf(logbuf, "not starting a local keepstore process because a volume (%s) uses AccessViaHosts\n", uuid)
2241                         return nil, nil
2242                 }
2243                 if !vol.ReadOnly && vol.Replication < configData.Cluster.Collections.DefaultReplication {
2244                         fmt.Fprintf(logbuf, "not starting a local keepstore process because a writable volume (%s) has replication less than Collections.DefaultReplication (%d < %d)\n", uuid, vol.Replication, configData.Cluster.Collections.DefaultReplication)
2245                         return nil, nil
2246                 }
2247         }
2248
2249         // Rather than have an alternate way to tell keepstore how
2250         // many buffers to use when starting it this way, we just
2251         // modify the cluster configuration that we feed it on stdin.
2252         configData.Cluster.API.MaxKeepBlobBuffers = configData.KeepBuffers
2253
2254         localaddr := localKeepstoreAddr()
2255         ln, err := net.Listen("tcp", net.JoinHostPort(localaddr, "0"))
2256         if err != nil {
2257                 return nil, err
2258         }
2259         _, port, err := net.SplitHostPort(ln.Addr().String())
2260         if err != nil {
2261                 ln.Close()
2262                 return nil, err
2263         }
2264         ln.Close()
2265         url := "http://" + net.JoinHostPort(localaddr, port)
2266
2267         fmt.Fprintf(logbuf, "starting keepstore on %s\n", url)
2268
2269         var confJSON bytes.Buffer
2270         err = json.NewEncoder(&confJSON).Encode(arvados.Config{
2271                 Clusters: map[string]arvados.Cluster{
2272                         configData.Cluster.ClusterID: *configData.Cluster,
2273                 },
2274         })
2275         if err != nil {
2276                 return nil, err
2277         }
2278         cmd := exec.Command("/proc/self/exe", "keepstore", "-config=-")
2279         if target, err := os.Readlink(cmd.Path); err == nil && strings.HasSuffix(target, ".test") {
2280                 // If we're a 'go test' process, running
2281                 // /proc/self/exe would start the test suite in a
2282                 // child process, which is not what we want.
2283                 cmd.Path, _ = exec.LookPath("go")
2284                 cmd.Args = append([]string{"go", "run", "../../cmd/arvados-server"}, cmd.Args[1:]...)
2285                 cmd.Env = os.Environ()
2286         }
2287         cmd.Stdin = &confJSON
2288         cmd.Stdout = logbuf
2289         cmd.Stderr = logbuf
2290         cmd.Env = append(cmd.Env,
2291                 "GOGC=10",
2292                 "ARVADOS_SERVICE_INTERNAL_URL="+url)
2293         err = cmd.Start()
2294         if err != nil {
2295                 return nil, fmt.Errorf("error starting keepstore process: %w", err)
2296         }
2297         cmdExited := false
2298         go func() {
2299                 cmd.Wait()
2300                 cmdExited = true
2301         }()
2302         ctx, cancel := context.WithDeadline(context.Background(), time.Now().Add(time.Second*10))
2303         defer cancel()
2304         poll := time.NewTicker(time.Second / 10)
2305         defer poll.Stop()
2306         client := http.Client{}
2307         for range poll.C {
2308                 testReq, err := http.NewRequestWithContext(ctx, "GET", url+"/_health/ping", nil)
2309                 testReq.Header.Set("Authorization", "Bearer "+configData.Cluster.ManagementToken)
2310                 if err != nil {
2311                         return nil, err
2312                 }
2313                 resp, err := client.Do(testReq)
2314                 if err == nil {
2315                         resp.Body.Close()
2316                         if resp.StatusCode == http.StatusOK {
2317                                 break
2318                         }
2319                 }
2320                 if cmdExited {
2321                         return nil, fmt.Errorf("keepstore child process exited")
2322                 }
2323                 if ctx.Err() != nil {
2324                         return nil, fmt.Errorf("timed out waiting for new keepstore process to report healthy")
2325                 }
2326         }
2327         os.Setenv("ARVADOS_KEEP_SERVICES", url)
2328         return cmd, nil
2329 }
2330
2331 // return current uid, gid, groups in a format suitable for logging:
2332 // "crunch-run process has uid=1234(arvados) gid=1234(arvados)
2333 // groups=1234(arvados),114(fuse)"
2334 func currentUserAndGroups() string {
2335         u, err := user.Current()
2336         if err != nil {
2337                 return fmt.Sprintf("error getting current user ID: %s", err)
2338         }
2339         s := fmt.Sprintf("crunch-run process has uid=%s(%s) gid=%s", u.Uid, u.Username, u.Gid)
2340         if g, err := user.LookupGroupId(u.Gid); err == nil {
2341                 s += fmt.Sprintf("(%s)", g.Name)
2342         }
2343         s += " groups="
2344         if gids, err := u.GroupIds(); err == nil {
2345                 for i, gid := range gids {
2346                         if i > 0 {
2347                                 s += ","
2348                         }
2349                         s += gid
2350                         if g, err := user.LookupGroupId(gid); err == nil {
2351                                 s += fmt.Sprintf("(%s)", g.Name)
2352                         }
2353                 }
2354         }
2355         return s
2356 }
2357
2358 // Return a suitable local interface address for a local keepstore
2359 // service. Currently this is the numerically lowest non-loopback ipv4
2360 // address assigned to a local interface that is not in any of the
2361 // link-local/vpn/loopback ranges 169.254/16, 100.64/10, or 127/8.
2362 func localKeepstoreAddr() string {
2363         var ips []net.IP
2364         // Ignore error (proceed with zero IPs)
2365         addrs, _ := processIPs(os.Getpid())
2366         for addr := range addrs {
2367                 ip := net.ParseIP(addr)
2368                 if ip == nil {
2369                         // invalid
2370                         continue
2371                 }
2372                 if ip.Mask(net.CIDRMask(8, 32)).Equal(net.IPv4(127, 0, 0, 0)) ||
2373                         ip.Mask(net.CIDRMask(10, 32)).Equal(net.IPv4(100, 64, 0, 0)) ||
2374                         ip.Mask(net.CIDRMask(16, 32)).Equal(net.IPv4(169, 254, 0, 0)) {
2375                         // unsuitable
2376                         continue
2377                 }
2378                 ips = append(ips, ip)
2379         }
2380         if len(ips) == 0 {
2381                 return "0.0.0.0"
2382         }
2383         sort.Slice(ips, func(ii, jj int) bool {
2384                 i, j := ips[ii], ips[jj]
2385                 if len(i) != len(j) {
2386                         return len(i) < len(j)
2387                 }
2388                 for x := range i {
2389                         if i[x] != j[x] {
2390                                 return i[x] < j[x]
2391                         }
2392                 }
2393                 return false
2394         })
2395         return ips[0].String()
2396 }
2397
2398 func (cr *ContainerRunner) loadPrices() {
2399         buf, err := os.ReadFile(filepath.Join(lockdir, pricesfile))
2400         if err != nil {
2401                 if !os.IsNotExist(err) {
2402                         cr.CrunchLog.Printf("loadPrices: read: %s", err)
2403                 }
2404                 return
2405         }
2406         var prices []cloud.InstancePrice
2407         err = json.Unmarshal(buf, &prices)
2408         if err != nil {
2409                 cr.CrunchLog.Printf("loadPrices: decode: %s", err)
2410                 return
2411         }
2412         cr.pricesLock.Lock()
2413         defer cr.pricesLock.Unlock()
2414         var lastKnown time.Time
2415         if len(cr.prices) > 0 {
2416                 lastKnown = cr.prices[0].StartTime
2417         }
2418         cr.prices = cloud.NormalizePriceHistory(append(prices, cr.prices...))
2419         for i := len(cr.prices) - 1; i >= 0; i-- {
2420                 price := cr.prices[i]
2421                 if price.StartTime.After(lastKnown) {
2422                         cr.CrunchLog.Printf("Instance price changed to %#.3g at %s", price.Price, price.StartTime.UTC())
2423                 }
2424         }
2425 }
2426
2427 func (cr *ContainerRunner) calculateCost(now time.Time) float64 {
2428         cr.pricesLock.Lock()
2429         defer cr.pricesLock.Unlock()
2430
2431         // First, make a "prices" slice with the real data as far back
2432         // as it goes, and (if needed) a "since the beginning of time"
2433         // placeholder containing a reasonable guess about what the
2434         // price was between cr.costStartTime and the earliest real
2435         // data point.
2436         prices := cr.prices
2437         if len(prices) == 0 {
2438                 // use price info in InstanceType record initially
2439                 // provided by cloud dispatcher
2440                 var p float64
2441                 var it arvados.InstanceType
2442                 if j := os.Getenv("InstanceType"); j != "" && json.Unmarshal([]byte(j), &it) == nil && it.Price > 0 {
2443                         p = it.Price
2444                 }
2445                 prices = []cloud.InstancePrice{{Price: p}}
2446         } else if prices[len(prices)-1].StartTime.After(cr.costStartTime) {
2447                 // guess earlier pricing was the same as the earliest
2448                 // price we know about
2449                 filler := prices[len(prices)-1]
2450                 filler.StartTime = time.Time{}
2451                 prices = append(prices, filler)
2452         }
2453
2454         // Now that our history of price changes goes back at least as
2455         // far as cr.costStartTime, add up the costs for each
2456         // interval.
2457         cost := 0.0
2458         spanEnd := now
2459         for _, ip := range prices {
2460                 spanStart := ip.StartTime
2461                 if spanStart.After(now) {
2462                         // pricing information from the future -- not
2463                         // expected from AWS, but possible in
2464                         // principle, and exercised by tests.
2465                         continue
2466                 }
2467                 last := false
2468                 if spanStart.Before(cr.costStartTime) {
2469                         spanStart = cr.costStartTime
2470                         last = true
2471                 }
2472                 cost += ip.Price * spanEnd.Sub(spanStart).Seconds() / 3600
2473                 if last {
2474                         break
2475                 }
2476                 spanEnd = spanStart
2477         }
2478
2479         return cost
2480 }
2481
2482 func (runner *ContainerRunner) handleSIGUSR2(sigchan chan os.Signal) {
2483         for range sigchan {
2484                 runner.loadPrices()
2485                 update := arvadosclient.Dict{
2486                         "select": []string{"uuid"},
2487                         "container": arvadosclient.Dict{
2488                                 "cost": runner.calculateCost(time.Now()),
2489                         },
2490                 }
2491                 runner.DispatcherArvClient.Update("containers", runner.Container.UUID, update, nil)
2492         }
2493 }